Hardy Spaces 9781107184541, 9781316882108

Citation preview

Appendix C Key Notions of Hilbert Spaces

In this chapter, every vector space is over the field C of complex numbers. For the properties shared by all Banach spaces, see Appendix D.

C.1 Scalar Products and Hilbert Spaces Let H be a vector space. A complex function (·, ·) = (·, ·)H on H × H is called a scalar product if it satisfies the following properties: (i) (ii) (iii) (iv)

x −→ (x, y) is a linear functional on H for any y ∈ H, (x, y) = (y, x) for every x, y ∈ H, (x, x) ≥ 0 for every x ∈ H, (x, x) = 0 ⇔ x = 0.

2 • Cauchy–Schwarz inequality. |(x, y)| < (x, x) · (y, y), except in the case where x and y are collinear (where equality holds in place of the inequality). 1/2 • Given a scalar product (·, ·), the function x −→ x = (x, x) is a norm on H. A vector space H equipped with a scalar product (·, ·) = (·, ·)H and with the associated norm is called a pre-Hilbert (or Hermitian) space; if it is complete (as a normed space, see Appendix D), it is said to be a Hilbert space.  2 2 • Example. H = L (Ω, μ) with ( f, g) = Ω f g dμ ( f, g ∈ L (Ω, μ)). In particular,



|x j |2 < ∞ l2 (J) = (x j ) j∈J : x j ∈ C,  j with (x, y) = j∈J x j y j .

In what follows, H always denotes a Hilbert space. 247 https://doi.org/10.1017/9781316882108.011 Published online by Cambridge University Press

248

Key Notions of Hilbert Spaces

C.2 Orthogonal Decompositions Let x, y ∈ H. An element x is said to be orthogonal to y (written x ⊥ y) if (x, y) = 0. Subspaces E, F ⊂ H are orthogonal (E ⊥ F) if x ⊥ y for every x ∈ E, y ∈ F. theorem (580–495 BCE). If x j ∈ H and x j ⊥ xk ( j  k), • The Pythagorean   then  n1 x j 2 = n1 x j 2 . • Corollary. A vector sum of closed and orthogonal subspaces is closed: if E, F ⊂ H are closed and E ⊥ F then E + F is closed (this is not necessarily the case for arbitrary E, F). • The orthogonal complement of a vector subspace E ⊂ H is E ⊥ = {y ∈ H : x ⊥ y ∀x ∈ E}. If E is closed, then E = (E ⊥ )⊥ and H = E + E ⊥ (often written as H = E ⊕ E ⊥ to highlight the orthogonality), hence every x ∈ H can be uniquely written in the form x = x + x where x ∈ E, x ∈ E ⊥ . The mapping PE : x −→ x is called the orthogonal projection onto E. Clearly PE is linear, with P2E = PE , and for every x ∈ H, PE x ≤ x. • Corollary. Let A ⊂ H. Then, spanH (A) = H ⇔ (x ⊥ A ⇒ x = 0). orthogonal series. Let x j ∈ H ( j = 1, 2, . . . ) and x j ⊥ xk • Convergence of an   ( j  k). The series j≥1 x j converges in H if and only if j x j 2 < ∞; in    this case, x2 = j x j 2 where x = j x j . An orthogonal series j x j converges unconditionally (if it converges): i.e. for any > 0 there exists a  finite set σ ⊂ N such that for every finite σ ⊃ σ , x − j∈σ x j  <

• Orthogonal decomposition. Let H j ⊂ H ( j = 1, 2, . . . ) and H j ⊥ Hk ( j  k). Then the closed linear hull of the family (H j ) is 



  x j : x j ∈ H j (∀ j) and x j 2 < ∞ . spanH H j : j = 1, 2, . . . = x = j

j≥1

This is denoted

 j≥1

⊕H j .

 • Orthogonal decomposition (continued). We have j≥1 ⊕H j = H if and only if (x ⊥ H j , ∀ j ⇒ x = 0), and if this is the case, then for every x ∈ H,



x= PH j x, PH j x ∈ H j , x2 = PH j x2 j

(Parseval’s identity).

https://doi.org/10.1017/9781316882108.011 Published online by Cambridge University Press

j

C.3 Orthogonal Bases

249

C.3 Orthogonal Bases A special case of the preceding decompositions is when H j is generated by a single vector e j  0; hence (e j ) is an orthogonal sequence, (e j , ek ) = 0, j  k. The sequence (e j ) is said to be complete in H if (x ⊥ e j , ∀ j ⇒ x = 0). In this case, for every x ∈ H there exists a unique convergent series of the form  j a j e j whose sum is x; indeed, a j e j = PH j x =

(x, e j ) e j, e j 2

thus ∀x ∈ H : x =

(x, e j ) j

e j 2

e j,

x2 =

|(x, e j )|2 j

e j 2

.

Such an orthogonal and complete sequence is called an orthogonal basis of H; if ∀ j we have e j  = 1, it is said to be an orthonormal basis. • The existence of an orthonormal basis (Gram–Schmidt orthogonalization theorem). Let (x j ) j≥1 ⊂ H be a “free” sequence (∀k, xk  Lin(x j : j  k)). Then there exists a unique orthonormal sequence (e j ) j≥1 satisfying the following properties. (i) For every n = 1, 2, . . . , Lin(e j : 1 ≤ j ≤ n) = Lin(x j : 1 ≤ j ≤ n) := Ln . (ii) For every j, (x j , e j ) > 0. The explicit formula is en =

xn − PLn−1 xn , xn − PLn−1 xn 

PLn−1 x =

n−1

(x, e j )e j

(∀x ∈ H).

j=1

• Corollary. In every separable Hilbert space there exists an orthonormal basis. • Example. Let μ be a finite Borel measure on T, such that supp(μ) is an infinite set. Then there exists a unique orthonormal basis (ϕk )k≥1 of trigonometric polynomials ϕk such that deg(ϕk ) = [k/2], k = 1, 2, . . . . (We apply the theorem to (xk )k≥1 = (1, eix , e−ix , e2ix , e−2ix , . . .); spanL2 (μ) (xk )k≥1 = L2 (μ) by § A.5 above.) The ϕk are called the orthogonal polynomials with respect to μ. • Corollary. All separable Hilbert spaces of the same dimension are unitarily isomorphic: if dim H1 = dim H2 (and the H j are separable) there exists a unitary (linear bijective isometric) U : H1 → H2 .

https://doi.org/10.1017/9781316882108.011 Published online by Cambridge University Press

250

Key Notions of Hilbert Spaces

C.4 The Riesz Representation Theorem Every linear continuous (bounded) functional ϕ on a Hilbert space H is of the form ϕ(x) = (x, y) (∀x ∈ H); such a y ∈ H is unique and we have ϕ = y.

https://doi.org/10.1017/9781316882108.011 Published online by Cambridge University Press

C A M B R I D G E S T U D I E S I N A DVA N C E D M AT H E M AT I C S 1 7 9 Editorial Board ´ S , W. F U LT O N , F. K I R WA N , B. BOLLOBA P. S A R N A K , B . S I M O N , B . T O TA R O

HARDY SPACES The theory of Hardy spaces is a cornerstone of modern analysis. It combines techniques from functional analysis, the theory of analytic functions, and Lesbesgue integration to create a powerful tool for many applications, pure and applied, from signal processing and Fourier analysis to maximum modulus principles and the Riemann zeta function. This book, aimed at beginning graduate students, introduces and develops the classical results on Hardy spaces and applies them to fundamental concrete problems in analysis. The results are illustrated with numerous solved exercises which also introduce subsidiary topics and recent developments. The reader’s understanding of the current state of the field, as well as its history, are further aided by engaging accounts of the key players and by the surveys of recent advances (with commented reference lists) that end each chapter. Such broad coverage makes this book the ideal source on Hardy spaces. Nikola¨ı Nikolski is Professor Emeritus at the Universit´e de Bordeaux working primarily in analysis and operator theory. He has been co-editor of four international journals and published numerous articles and research monographs. He has also supervised some 30 PhD students, including three Salem Prize winners. Professor Nikolski was elected Fellow of the AMS in 2013 and received the Prix Amp`ere of the French Academy of Sciences in 2010.

Published online by Cambridge University Press

CAMBRIDGE STUDIES IN ADVANCED MATHEMATICS Editorial Board B. Bollob´as, W. Fulton, F. Kirwan, P. Sarnak, B. Simon, B. Totaro All the titles listed below can be obtained from good booksellers or from Cambridge University Press. For a complete series listing, visit www.cambridge.org/mathematics. Already Published 140 R. Pemantle & M. C. Wilson Analytic Combinatorics in Several Variables 141 B. Branner & N. Fagella Quasiconformal Surgery in Holomorphic Dynamics 142 R. M. Dudley Uniform Central Limit Theorems (2nd Edition) 143 T. Leinster Basic Category Theory 144 I. Arzhantsev, U. Derenthal, J. Hausen & A. Laface Cox Rings 145 M. Viana Lectures on Lyapunov Exponents 146 J.-H. Evertse & K. Gy˝ory Unit Equations in Diophantine Number Theory 147 A. Prasad Representation Theory 148 S. R. Garcia, J. Mashreghi & W. T. Ross Introduction to Model Spaces and Their Operators 149 C. Godsil & K. Meagher Erd˝os–Ko–Rado Theorems: Algebraic Approaches 150 P. Mattila Fourier Analysis and Hausdorff Dimension 151 M. Viana & K. Oliveira Foundations of Ergodic Theory 152 V. I. Paulsen & M. Raghupathi An Introduction to the Theory of Reproducing Kernel Hilbert Spaces 153 R. Beals & R. Wong Special Functions and Orthogonal Polynomials 154 V. Jurdjevic Optimal Control and Geometry: Integrable Systems 155 G. Pisier Martingales in Banach Spaces 156 C. T. C. Wall Differential Topology 157 J. C. Robinson, J. L. Rodrigo & W. Sadowski The Three-Dimensional Navier–Stokes Equations 158 D. Huybrechts Lectures on K3 Surfaces 159 H. Matsumoto & S. Taniguchi Stochastic Analysis 160 A. Borodin & G. Olshanski Representations of the Infinite Symmetric Group 161 P. Webb Finite Group Representations for the Pure Mathematician 162 C. J. Bishop & Y. Peres Fractals in Probability and Analysis 163 A. Bovier Gaussian Processes on Trees 164 P. Schneider Galois Representations and (ϕ, Γ)-Modules 165 P. Gille & T. Szamuely Central Simple Algebras and Galois Cohomology (2nd Edition) 166 D. Li & H. Queffelec Introduction to Banach Spaces, I 167 D. Li & H. Queffelec Introduction to Banach Spaces, II 168 J. Carlson, S. M¨uller-Stach & C. Peters Period Mappings and Period Domains (2nd Edition) 169 J. M. Landsberg Geometry and Complexity Theory 170 J. S. Milne Algebraic Groups 171 J. Gough & J. Kupsch Quantum Fields and Processes 172 T. Ceccherini-Silberstein, F. Scarabotti & F. Tolli Discrete Harmonic Analysis 173 P. Garrett Modern Analysis of Automorphic Forms by Example, I 174 P. Garrett Modern Analysis of Automorphic Forms by Example, II 175 G. Navarro Character Theory and the McKay Conjecture 176 P. Fleig, H. P. A. Gustafsson, A. Kleinschmidt & D. Persson Eisenstein Series and Automorphic Representations 177 E. Peterson Formal Geometry and Bordism Operators 178 A. Ogus Lectures on Logarithmic Algebraic Geometry 179 N. Nikolski Hardy Spaces

Published online by Cambridge University Press

Hardy Spaces N I K O L A ¨I N I K O L S K I Universit´e de Bordeaux

Published online by Cambridge University Press

University Printing House, Cambridge CB2 8BS, United Kingdom One Liberty Plaza, 20th Floor, New York, NY 10006, USA 477 Williamstown Road, Port Melbourne, VIC 3207, Australia 314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India 79 Anson Road, #06–04/06, Singapore 079906 Cambridge University Press is part of the University of Cambridge. It furthers the University’s mission by disseminating knowledge in the pursuit of education, learning, and research at the highest international levels of excellence. www.cambridge.org Information on this title: www.cambridge.org/9781107184541 DOI: 10.1017/9781316882108 ´ ements d’analyse avanc´ee: Originally published in French as El´ ´ 1. Espaces de Hardy by Belin, 2012. © Editions Belin, 2012 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in English by Cambridge University Press 2019 English translation © Cambridge University Press 2019 Printed in the United Kingdom by TJ International Ltd. Padstow Cornwall A catalogue record for this publication is available from the British Library. Library of Congress Cataloging-in-Publication Data Names: Nikolski, N. K. (Nikolai Kapitonovich), author. Title: Hardy spaces : elements of advanced analysis / Nikolai Nikolski (Universite de Bordeaux). Other titles: Elements d’analyse avancee. 1, Espaces de Hardy. English | Espaces de Hardy Description: Cambridge ; New York, NY : Cambridge University Press, 2019. | Series: Cambridge studies in advanced mathematics ; 179 | Originally published in French: Elements d’analyse avancee : 1, Espaces de Hardy (Paris : Editions Belin, 2012). | First English translation. | Includes bibliographical references and index. Identifiers: LCCN 2018049103 | ISBN 9781107184541 (hardback : alk. paper) Subjects: LCSH: Hardy spaces. | Functions of complex variables. | Holomorphic functions. Classification: LCC QA331.7 .N5513 2019 | DDC 515/.98–dc23 LC record available at https://lccn.loc.gov/2018049103 ISBN 978-1-107-18454-1 Hardback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication and does not guarantee that any content on such websites is, or will remain, accurate or appropriate. Every effort has been made to secure necessary permissions to reproduce copyright material in this work, though in some cases it has proved impossible to trace copyright holders. If any omissions are brought to our notice, we will be happy to include appropriate acknowledgements on reprinting.

Published online by Cambridge University Press

Less is more Robert Browning, “Andrea del Sarto,” 1855

Published online by Cambridge University Press

Published online by Cambridge University Press

Contents

Preface Acknowledgments for the French Edition List of Biographies List of Figures The Origins of the Subject 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8

1.9 2 2.1 2.2 2.3 2.4

page xiii xv xvi xvii 1

The space H 2 (T): An Archetypal Invariant Subspace Notation and Terminology of Operators Reducing Subspaces of the Bilateral Shift Mz Non-reducing Subspaces of the Bilateral Shift Mz 1.3.1 H p (T) Spaces Beurling “Inner Functions” H 2 (μ) Spaces and the Riesz Brothers’ Theorem 1.5.1 Elementary Proof of Theorem 1.5.4 (Øksendal, 1971) The Past and the Future: The Prediction Problem Inner–Outer Factorization and Szeg˝o’s Infimum Exercises 1.8.1 The Wold–Kolmogorov Decomposition 1.8.2 The Shift Operator Mz on L2 (T, μ) 1.8.3 Inner and Outer Functions Notes and Remarks

5 5 6 8 13 14 16 18 20 25 29 29 30 31 34

The H p (D) Classes: Canonical Factorization and First Applications Fej´er and Poisson Means Definition of H p (D): Identification of H p (D) and H p (T) Jensen’s Formula and Jensen’s Inequality: log | f | ∈ L1 (T) Blaschke Products

37 37 40 42 46

vii Published online by Cambridge University Press

viii

2.5 2.6 2.7

2.8

2.9 3 3.1 3.2

3.3 3.4

3.5

3.6 4 4.1

Contents

Fatou’s Theorem (Non-tangential Boundary Values) The Smirnov Canonical Factorization Applications: Szeg˝o Infimum, Weighted Polynomial Approximations, Invariant Subspaces of L p (T) 2.7.1 Cyclic Vectors of the Shift Operator Mz 2.7.2 Weighted Density of Polynomials Pa Exercises 2.8.1 Invariant Subspaces of L p (T, μ) 2.8.2 Factorization on the H p Scale, 0 < p < ∞ 2.8.3 The Hilbert and Hardy Inequalities 2.8.4 Harmonic Conjugates and the Riesz Projection (1927), Following Calder´on (1950) 2.8.5 The Kolmogorov Weak Type Inequality 2.8.6 The Littlewood Subordination Principle (1925) Notes and Remarks

49 53 57 58 59 60 60 63 65 68 73 74 77

The Smirnov Class D and the Maximum Principle Calculus of Outer Functions 3.1.1 Properties of Outer Functions Calculus of Inner Functions: The Spectrum 3.2.1 Properties of the Divisors, GCDs, and LCMs 3.2.2 Logarithmic Residues The Nevanlinna (N) and Smirnov (D) Classes 3.3.1 A Few Properties of N and D, by Smirnov (1932) The Generalized Phragm´en–Lindel¨of Principle 3.4.1 The Spaces N and D: Conformally Invariant Versions 3.4.2 Generalized Phragm´en–Lindel¨of Principle 3.4.3 Classical Examples Exercises 3.5.1 An Improvement of Liouville’s Theorem 3.5.2 The Case of a Strip (Phragm´en and Lindel¨of, 1908) 3.5.3 An Inner Function Which Becomes Outer on a Subdomain 3.5.4 Division by a Singular Function with a Point Measure Notes and Remarks

82 82 83 86 87 91 92 94 96 96 97 99 100 100 102 102 103 104

An Introduction to Weighted Fourier Analysis Generalized Fourier Series 4.1.1 Minimal Sequences 4.1.2 Bases

106 108 109 111

Published online by Cambridge University Press

Contents

4.2

ix

Skew Projections 4.2.1 Properties of PLM 4.3 The Angle Between the Past and the Future 4.3.1 Properties of the Angle 4.4 The Case of the Exponentials: A Reduction to P+ 4.5 The Hilbert Operator: The Classical Case of L2 (T) 4.6 Exponential Bases in L2 (T, μ) 4.7 Prediction and Hankel Operators 4.7.1 Strongly Regular Processes 4.7.2 Angular Operators and Hankel Operators 4.8 b(X) Versus ub(X) 4.8.1 Proof of Theorem 4.8.2 4.8.2 Gram Matrices 4.9 Exercises 4.9.1 Criterion of Linear Dependence of Exponentials 4.9.2 Multipliers Versus Bases 4.9.3 Projections on a Hilbert Space 4.9.4 The Sharpness of the McCarthy–Schwartz Inequality 4.10 Notes and Remarks

114 114 115 115 116 119 120 125 126 127 129 135 135 138 138 138 141 141 145

5 5.1

150 150 152 153 156 156 157 158 160 161 164 166 167 169 170 170 171 175 176

5.2 5.3

5.4

5.5 5.6

5.7

Harmonic Analysis and Stationary Filtering The Language of Linear Filters 5.1.1 The Fourier Transform and the Frequency Domain Characterization of Stationary Filters What Can Filtering Do? 5.3.1 A Bit More Terminology for Filters 5.3.2 Some Typical Problems in Filtering Synthesis of Causal Filters 5.4.1 Filters of Optimal “Signal to Noise Ratio” 5.4.2 Frequency Response on a Very Thin Band 5.4.3 Helson Sets: Arbitrary Frequency Response on σ ⊂ T 5.4.4 Causal Recursive Filters Inverse Problem: “Can One Hear the Shape of a Drum?” 5.5.1 Moving Averages of a Signal Exercises 5.6.1 Identification of Filters: Moving Averages 5.6.2 The Non-equality Ca (D)  Wa (D) 5.6.3 Helson Sets in the Disk D (Vinogradov, 1965) Notes and Remarks

Published online by Cambridge University Press

x

6 6.1

6.2 6.3

6.4 6.5 6.6

6.7

6.8

Contents

The Riemann Hypothesis, Dilations, and H 2 in the Hilbert Multi-disk The Euler ζ Function and the Riemann Hypothesis (RH) 6.1.1 Prime Number Decomposition (Euclid, c. 300 BCE; Gauss, 1801) 6.1.2 The Euler Infinite Product 6.1.3 The Riemann Hypothesis (RH), 1859 An Approximation Implying the Riemann Hypothesis H 2 (C+ ) and the “Weak Paley–Wiener Theorem” 6.3.1 A Unitary Mapping of L2 (T) onto L2 (R) 6.3.2 Fourier Transforms and the “Weak Paley–Wiener Theorem” 6.3.3 The Mellin Transform and the Group of Dilations 6.3.4 Completeness of the Characters, the Translations, and/or the Dilations The Nyman Theorem The Distance Function and Zero-free Disks of ζ 6.5.1 The Distance Function Completeness of Dilations and the Hilbert Multi-disk 6.6.1 The Wintner–Beurling Problem 6.6.2 Change of Orthonormal Basis: The Semigroup T = (T n ) on H02 6.6.3 The Reproduction of Variables and the Bohr Transform 2 ∞ 6.6.4 The Hilbert Multi-disk D∞ 2 and the Space H (D2 ) 6.6.5 A Few Initial Observations 6.6.6 Cyclic Polynomials 6.6.7 Other Classes of (T n )-cyclic Functions of H02 Exercises 6.7.1 Multipliers of the Space H 2 (D∞ 2 ) 6.7.2 Orthogonal Dilations 6.7.3 Asymptotics of ka  pp as a → 0 6.7.4 Particular Features of the Multi-disk D∞ 2 6.7.5 A Few Cyclic Functions in H 2 (D∞ 2 ) 6.7.6 A Function (Dn )-cyclic in L2 (0, 1) (Kozlov, 1950; Akhiezer, 1965) Notes and Remarks

187 190 190 191 195 197 199 199 202 204 206 207 208 209 211 211 214 214 216 217 220 221 223 223 223 224 224 225 227 227

Appendix A

Key Notions of Integration

233

Appendix B

Key Notions of Complex Analysis

243

Appendix C

Key Notions of Hilbert Spaces

247

Published online by Cambridge University Press

Contents

xi

Appendix D

Key Notions of Banach spaces

251

Appendix E

Key Notions of Linear Operators

254

References Notation Index

Published online by Cambridge University Press

259 268 270

Published online by Cambridge University Press

Preface

The introduction to Hardy spaces proposed in this book covers the basic techniques of modern analysis, conceived and developed at the beginning of the twentieth century over a very short period (a kind of “Silver Age” for mathematical analysis; Exercise 1: which was the “Golden Age”?), by a talented group of mathematical geniuses including Henri Lebesgue, Frigyes Riesz, G. H. Hardy, Andrey Kolmogorov, and Norbert Wiener. Over time, this cluster of ideas became the source of extremely powerful techniques for a variety of applications: from Fourier series to the Wiener theory of stationary filtering, not to mention the Euler ζ function and the Riemann hypothesis. The contents of this text correspond to a course at the “Master 2” level given several times during the years 1990–2010 at the University of Bordeaux 1, and represent an introduction and invitation to the entire domain of modern analysis. The book is devoted to a multi-faceted subject: it involves harmonic analysis (since it concerns a unitary representation of the group Z), but also complex analysis (as we restrict ourselves most often to the semigroup Z+ ), the theory of operators (by the nature of the representation, but also by a hidden universality that we will explore in future volumes), as well as the theme of signals and filtering, with a bit of number theory thrown in. It is for this superposition of major disciplines of mathematics (more a “roundabout” than a “crossroads”) that the subject can be described as “classical” (“classical”  “old-fashioned”!). The conjunction between the different facets of the subject is most fruitful and successful in the Hilbert framework of the spaces L2 (T, μ); this is why we have developed the theory, and its applications, principally in the space H 2 (which is also closely linked with H 1 and H ∞ ), whereas the other H p spaces appear only occasionally. The prerequisites are a standard course in integration and functional analysis (or in Hilbert/Banach spaces) along with a few notions of complex analysis. A summary/reminder of all the necessary information (as well as xiii https://doi.org/10.1017/9781316882108.001 Published online by Cambridge University Press

xiv

Preface

certain notations) are gathered in the appendices at the end of the book. Within the text, we include a large number of historical details – on the subjects developed, their founders, and the diverse circumstances of their creation. We hope that this will help the reader to better understand Hardy spaces, along with the dramaturgy of mathematics (and mathematical life).1 Each chapter contains exercises and their solutions (75 in total) at different levels: to use a musical analogy from Glazman and Lyubich (1969), from exercises on open strings up to virtuoso pieces using double harmonics (“double flageolet tones”). Each chapter concludes with a section entitled “Notes and Remarks” which discusses the history of the main subjects of the chapter, certain recent results, and (at times) the open questions; this discussion is sometimes addressed to non-novice readers. The reader will rapidly become aware that this text contains only a few elementary aspects of the techniques of harmonic analysis, linked particularly with an approach to Hardy spaces via complex analysis. Even if at times we delve into quite refined questions of analysis (such as the geometry of finite bases, in Chapter 4), our text is not meant to be a research monograph, but more a source of basic knowledge. This is why “less is more.” Nonetheless, in principle, students reaching the end of the book should be capable of tackling independent research projects (the author can affirm this from experience). For such an endeavor they will need the aid of experts, but this can be found in the dozens of existing monographs devoted to Hardy spaces and the “hard analysis” that was developed around them. Several are mentioned at the end of Chapter 1, in the section Notes and Remarks 1.9. Good luck!

1

The biographical details – which, given the technical and financial constraints, are sketched here at best – are drawn from various sources, notably the MacTutor website of the University of St Andrews (Scotland), www-history.mcs.st-and.ac.uk/, and the free encyclopedia Wikipedia, https://en.wikipedia.org/wiki/.

https://doi.org/10.1017/9781316882108.001 Published online by Cambridge University Press

Acknowledgments

Acknowledgements for the French Edition This book could not have seen the light of day without the generous and wide´ Charpentier, whose enthusiasm ranging aid of my colleague at Bordeaux, Eric and availability supported me at several difficult points in the editing. I am also grateful to my colleagues from Saint Petersburg: Anton Baranov (who read a preliminary version and detected several “bugs”), and Andre¨ı Khrabrov (who mastered all the “TEXnical” problems). ´ I thank Editions Belin (and especially the editor responsible for this work, C. Counillon) for accepting my project and seeing it through to fruition. And of course, I owe an eternal debt towards my young family for their infinite patience (and many sacrifices).

Acknowledgements for the English Edition The author warmly thanks the translators Dani`ele Gibbons and Greg Gibbons for their high-quality job, for attention to all shades of meaning of the French text, and for a friendly collaboration at all stages of the work. The author is also sincerely grateful to CUP for including the book in this prestigious series, and to the whole CUP editorial team for their highly professional preparation of the manuscript and for their patience during his numerous slowdowns due to many other projects.

xv Published online by Cambridge University Press

Biographies

G. H. Hardy Norbert Wiener Henry Helson Arne Beurling Frigyes Riesz Marcel Riesz Andrey N. Kolmogorov Vladimir I. Smirnov G´abor Szeg˝o Johan Jensen Wilhelm Blaschke Pierre Fatou Donald Newman David Hilbert J. E. Littlewood Rolf Nevanlinna Lars Edvard Phragm´en Ernst Leonard Lindel¨of

page

2 6 11 14 18 18 22 26 28 45 48 49 64 67 76 93 98 98

Joseph Liouville Joseph Fourier Stefan Banach Hermann Hankel Jacob Schwartz Jørgen Gram Walter Rudin E. T. Whittaker Vladimir A. Kotelnikov Kinnosuke Ogura Claude E. Shannon Leonhard Euler Bernhard Riemann Raymond Paley Robert Hjalmar Mellin Aurel Wintner Henri Lebesgue

xvi Published online by Cambridge University Press

page

101 107 111 128 131 136 162 177 179 182 184 188 196 202 205 212 235

Figures

G. H. Hardy (© DR)

page

2

Trinity College, Cambridge (pinggr / iStock / Getty Images Plus)

2

The first page of Hardy’s paper defining “Hardy classes” (from Hardy (1915))

3

Norbert Wiener (Bettmann / Bettmann / Getty Images)

6

Massachusetts Institute of Technology (Education Images / Universal Images Group / Getty Images)

7

Andrey N. Kolmogorov (© DR)

22

Lomonosov Moscow State University (bladerunner7 / iStock / Getty Images Plus)

23

Vladimir I. Smirnov (MacTutor History of Mathematics Archive: www-history.mcs.st-and.ac.uk)

26

University of Saint Petersburg (TanyaSv / iStock / Getty Images Plus)

26

G´abor Szeg˝o (MacTutor History of Mathematics Archive)

28

Johan Jensen

45

Pierre Fatou

49

Stolz angle at the point ζ on the unit circle

50

David Hilbert (MacTutor History of Mathematics Archive)

67

J. E. Littlewood (MacTutor History of Mathematics Archive)

76

The arc Δ in the set Uζ , where Θ is separated from zero

90

Rolf Nevanlinna (MacTutor History of Mathematics Archive)

93

Joseph Liouville (© Archives Belin)

101

Joseph Fourier (© Archives Belin)

107

xvii Published online by Cambridge University Press

xviii

List of Figures

Fourier’s Th´eorie analytique de la chaleur (1822) (© Archives Belin)

108

Stefan Banach (© DR)

111

Per Enflo receiving a live goose from Stanisław Mazur in 1972

112

The “Scottish Book” (PlWiki, uploaded by Stako / CC-BY-SA-3.0: https://en.wikipedia.org/wiki/Scottish Book)

112

The shaded domain Ω

123

Hermann Hankel (© Archives Belin)

128

Littlewood’s crocodile (from Littlewood (1953))

175

E. T. Whittaker (MacTutor History of Mathematics Archive)

177

Vladimir A. Kotelnikov (Presidential Press and Information Office / Пресс-служба Президента России / www.kremlin.ru/CC-BY-4.0 / https://en.wikipedia.org/wiki/Vladimir Kotelnikov)

179

Letter from the journal Electricity rejecting Kotelnikov’s pioneering paper 181 The Marfino “sharashka,” a Soviet Gulag research laboratory near Moscow

181

Kinnosuke Ogura (from Butzer et al. (2011), reprinted by permission of Springer Nature) 182 Claude E. Shannon (Keystone / Stringer / Hulton Archive / Getty Images)

184

Leonhard Euler (Bettmann / Bettmann / Getty Images)

188

Bernhard Riemann (© Archives Belin)

196

A self-caricature by Lewis Carroll (Culture Club / Hulton Archive / Getty Images)

197

Raymond Paley (MacTutor History of Mathematics Archive)

202

A Paley graph

203

The Paley–Zygmund inequality

203

A shifted non-Euclidean disk free of zeros of the ζ function

210

Aurel Wintner during a seminar at the Niels Bohr Institute in Copenhagen in 1930 (Science & Society Picture Library / SSPL / Getty Images)

212

Henri Lebesgue (© DR)

235

Published online by Cambridge University Press

The Origins of the Subject

Prehistory. Cauchy – Fourier – Poisson – Weierstrass – Stieltjes – Fatou – Lebesgue – Hilbert – Parseval – Jensen. History. Lebesgue – Hardy – Luzin – Privalov – Schur – the Riesz brothers – Szeg˝o – Nevanlinna – Smirnov – Littlewood – Kolmogorov – Paley – Wiener – Zygmund. Legacy/Continuation. Stein – Fefferman – de Branges – Helson – Kahane – Garnett – Gamelin – Carleson – Sarason – Havin – Douglas – Arveson – Sz.Nagy – Foias – Fuhrmann – Lax – Phillips – Lacey, etc. The birth of Hardy spaces dates back to the year 1915, at Cambridge University. At the time, it went virtually unnoticed. Admittedly, the year 1915 can be considered as “unremarkable” only for their creator, the British mathematician G. H. Hardy (1877–1947). Sure enough, as usual, he had published a dozen (!) articles and research notes, but apparently no salient result emerged from his efforts that year, with one exception – if we equate a definition with a result.

1 https://doi.org/10.1017/9781316882108.002 Published online by Cambridge University Press

2

The Origins of the Subject

Godfrey Harold (G. H.) Hardy (1877–1947) was one of the founding fathers of modern “hard” analysis, and the author of several fundamental ideas that transformed such disciplines as Diophantine analysis, Tauberian theory, the summation of divergent series, Fourier series, the distribution of prime numbers, and the theory of the Euler ζ function. David Hilbert called him “the best mathematician in England.” Several theorems and mathematical creations are named after Hardy. His book A Mathematician’s Apology (1940) is a masterpiece on the philosophy and psychology of a mathematician. His remarkable essay “Orders of infinity: The ‘Infinit¨arcalc¨ul’ of Paul Du Bois-Reymond” (1910) inspired a chapter in Bourbaki’s treatise. He was a friend of the novelist and scientist C. P. Snow and a co-author with Littlewood, Ramanujan, Titchmarsh, Ingham, Landau, and Marcel Riesz.

Trinity College, Cambridge.

Specifically, in part of a short nine-page article published in the 1915 Proceedings of the London Mathematical Society, Hardy defined a family of spaces (“function classes”) of holomorphic functions. At the time, the event was barely noticed: either by the general public (preoccupied by the

https://doi.org/10.1017/9781316882108.002 Published online by Cambridge University Press

The Origins of the Subject

3

The first page of Hardy’s nine-page paper of 1915 defining “Hardy classes.” Who could have prophesied that this acorn would grow into such a mighty oak?

First World War), or by the scientific world (1915 was above all the year of Einstein’s General Relativity, as well as Wegener’s theory of Pangaea), or even by mathematicians. Nevertheless, it was a turning point for a number of disciplines linked to mathematical analysis: complex analysis (then flourishing), harmonic analysis, signal processing, and in particular several theories nonexistent at the time, but crucial today – the theory of operators, optimal control, diffusion theory, random processes. Later on in his career, Hardy himself returned several times to the theory of the spaces he had defined in 1915, which, at first glance, seemed to be merely an auxiliary tool. However, for its transformation into an indispensable, extremely powerful technique of analysis and for the majority of its applications, we are highly indebted to the efforts of the “Golden Team” of analysts of that time (such as Schur, Marcel Riesz, Frigyes Riesz, Szeg˝o, Nevanlinna, Luzin, Privalov, Smirnov, Kolmogorov, Paley, Wiener, Zygmund), and to their equally brilliant successors (such as Beurling, Stein, Fefferman, de Branges, Helson, Carleson, Kahane, Garnett, Gamelin, Sarason, Havin, Douglas, Sz.Nagy, Foias, Fuhrmann, Lax, Phillips).

https://doi.org/10.1017/9781316882108.002 Published online by Cambridge University Press

4

The Origins of the Subject

The explanation for its success can perhaps be summed up in just a few points: (1) the dynamics of the Hardy space einx H 2 , n ∈ Z, generates an orthonormal basis einx ∈ einx H 2  ei(n+1)x H 2 in the Lebesgue space L2 (−π, π); (2) the space H 2 is the “analytic half” of L2 (−π, π); (3) in H 2 , there is a property of factorization into elementary factors, similar to that of polynomials (in a sense, H 2 is a “factorial ring”). First of course come the definition and the basic properties. A remark for the experts: the current dominant approach to Hardy spaces is via real harmonic analysis (maximal functions, Hilbert transforms, etc.); thus it is unnecessary to differentiate between H 2 and H p , p  2, or between the groups where the space is defined (T, Tn , R, Rn , etc., and even without any group structure). In this book, I follow a combination of the “genetic” approach based on analysis of a single complex variable, and the spectral analysis of a unitary representation of Z. Why this choice? It is indeed the most elementary and direct route to obtain all the results of the theory needed for applications. Let us add that, so far, the true value of the powerful methods of real variables remains purely theoretical. As soon as we are faced with practical applications of Hardy spaces, we use the complex presentation and its techniques – beginning with signal processing and operator theory, and then H ∞ optimal control and diffusion theory, or even stochastic processes or the Euler ζ function. Our work is especially concerned with the spaces H 2 , H 1 , and H ∞ .

The memorable events of 1915 • Einstein’s theory of General Relativity. • Wegener’s theory of Pangaea. • The use of chemical weapons by Germany on a massive scale (Second Battle of Ypres). • The Mexican Revolution. • The birth of Paul Tibbets (future pilot in the US Air Force, to be assigned the task of dropping the first atomic bomb on Hiroshima on August 6, 1945). • The thesis of Nikolai Luzin (future founder of the Moscow school of analysis), written in Paris and defended in Moscow. p • G. H. Hardy’s definition of H spaces.

https://doi.org/10.1017/9781316882108.002 Published online by Cambridge University Press

1 The space H 2 (T): An Archetypal Invariant Subspace

Topics. Lebesgue spaces L p (T, μ), Hardy spaces H p (T), lattice of invariant subspaces, the shift operator (reducing subspaces – Wiener’s theorem – and invariant subspaces – Helson’s theorem), uniqueness theorem, and inner and outer functions. In this chapter we mainly work in the context of the Hilbert spaces L2 (T, μ), L2 (T), H 2 (T); the other H p appear occasionally.

1.1 Notation and Terminology of Operators Let H be a Hilbert space (always over the field of complex numbers C) and let T : H → H be a bounded linear operator on H. The space (the algebra) of operators on H is denoted L(H). Let E ⊂ H be a subspace of H (= closed linear subspace). E is said to be invariant for T ∈ L(H) if x ∈ E ⇒ Tx ∈ E (in short, T E ⊂ E). The set Lat(T ) of invariant subspaces is a lattice with respect to the operations ∩ and span (= closed linear hull). If T is a family of operators on H, we set Lat(T ) =  T ∈T Lat(T ). In the particular case of T = {T, T ∗ }, where T ∈ L(H) and T ∗ is the adjoint operator of T (see Appendix E), a subspace E ∈ Lat(T, T ∗ ) is said to be reducing. The goal of this section is to describe the lattice Lat(Mz ) where Mz is the operator of multiplication by an “independent variable” in the space L2 (T, μ), 5 https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

6

The space H 2 (T): An Archetypal Invariant Subspace

with μ a finite Borel measure on the circle T = {ζ ∈ C : |ζ| = 1}, Mz f = z f (z), z ∈ T. The operator Mz is called the bilateral shift operator.

1.2 Reducing Subspaces of the Bilateral Shift M z In the years 1920–1930, Norbert Wiener developed the mathematical theory of stationary filters. Since the tools he needed could not be found in the Analysis of the time, he created them himself, thus profoundly enriching harmonic analysis and spectral theory.

Norbert Wiener (1894–1964) was an American mathematician (MIT: Massachusetts Institute of Technology), creator of cybernetics (1948) and communication theory (co-founded with Kotelnikov and Shannon). He also created the theories of stochastic processes and generalized harmonic analysis (1930, the Wiener measure and Brownian motion), Tauberian theory, and also, independently of Stefan Banach, invented Banach spaces (1923). He authored innovative works in mathematical physics, in potential theory and the optimal prediction of random processes (with applications to the automatic correction of the firing of anti-aircraft guns, shared with Kolmogorov). An admirer of Leibniz, Lebesgue, and Hadamard, Norbert Wiener was one of the geniuses of the twentieth century, who revolutionized mathematics and science. The reader can find a remarkable overview of Wiener’s scientific impact (as well as a biographical article by Norman Levinson) in vol. 72, issue 1-ii (1966) of the Bulletin of the American Mathematical Society. Having received his Bachelor’s degree at the age of 14, Wiener followed a Master’s program in zoology at Harvard, in philosophy at Cornell, and then in mathematics at Harvard. After submitting his thesis in 1912 (at the age of 17), he came to Europe for post-doctoral studies. Upon his return to the USA, Wiener

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.2 Reducing Subspaces of the Bilateral Shift Mz

7

was denied a position at Harvard because of the anti-Semitic atmosphere of the establishment (George Birkhoff is often cited as one of his principal opponents, behind the scenes). Unlike other top-level scientists, Wiener was not invited to participate Massachusetts Institute of Technology. in the Manhattan Project. A confirmed pacifist, he systematically refused all government financing of his research after the Second World War and never participated in military projects. In particular, for filtering theory, Wiener needed to solve the problem of the recognition (identification) of filters (see the details below in Chapter 5). As a first step, he proved the following theorem (in the case where μ = m, the normalized Lebesgue measure on the circle T; 80 years later, we prove it in a somewhat more general form). Theorem 1.2.1 (Wiener, 1932) Let μ be a positive Borel measure in C with compact support and E a (closed) subspace of L2 (μ). The following assertions are equivalent. (1) E ∈ Lat(Mz , Mz∗ ). (2) There exists a Borel set A ⊂ C such that E = χA L2 (μ) = { f ∈ L2 (μ) : f = 0 μ-a.e. on the complement A = C \ A}. The set A in (2) is unique modulo μ: χA L2 (μ) = χB L2 (μ) if and only if χA = χB μ-a.e., i.e. if and only if μ(A B) = 0, where A B = (A \ B) ∪ (B \ A) is the symmetric difference. Proof First observe that Mz∗ = Mz and 12 (z + z) = X, 2i1 (z − z) = Y imply that a subspace E is reducing for Mz if and only if, for every polynomial p = p(X, Y), we have p · E ⊂ E. Let P denote the set of polynomials in X and Y. Let us show (1) ⇒ (2). Let f ∈ E and g ∈ E ⊥ = {g ∈ L2 (μ) : (h, g) = 0, ∀h ∈ E} (orthogonal complement of E). Then  0 = (p f, g) = p f g dμ, ∀p ∈ P.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

8

The space H 2 (T): An Archetypal Invariant Subspace

Since P is dense in the space C(supp(μ)) of continuous functions on a compact set supp(μ) (Weierstrass’s theorem), we obtain f g dμ = 0 (the null measure), hence f g = 0 μ-a.e. Then, as L2 (μ) is separable, so is E ⊥ . By taking a sequence (gn ) dense in E ⊥ , we set  Z(gn ), Z(gn ) = {z : gn (z) = 0}. A= n

(More rigorously, we define Z(gn ) by choosing a measurable representative in the equivalence class gn of L2 (μ); another choice of representative would lead to a set A differing from A only by a negligible set, hence χA = χA in the space L2 (μ).) We obtain, for any f ∈ E and every n, f gn = 0 μ-a.e., and thus  f = 0 a.e. on the set n Z(gn )c = Ac . This means that f ∈ χA L2 (μ), and hence E ⊂ χA L2 (μ). Conversely, if f ∈ χA L2 (μ), then (clearly) f = 0 μ-a.e. on Ac . Since gn = 0 on A, we have f gn = 0 μ-a.e., thus ( f, gn ) = 0, ∀n. By the density of (gn ) in E ⊥ , we obtain f ⊥ E ⊥ , hence f ∈ E. The two inclusions give E = χA L2 (μ). The implication (2) ⇒ (1) is evident. For the uniqueness, the equality χA L2 (μ) = χB L2 (μ) implies χA ∈ χB L2 (μ), thus χA = 0 a.e. on Bc , meaning that A ⊂ B up to a μ-negligible set (i.e., μ(A \ B) = 0). Similarly, μ(B \ A) = 0, which completes the proof. 

1.3 Non-reducing Subspaces of the Bilateral Shift M z In order to catalog the non-reducing subspaces of Mz , we use two related (but not coincident) orthogonal decompositions. The first is given by Lemma 1.3.1 below and concerns an invariant subspace of an arbitrary operator. The second is the Radon–Nikodym decomposition (see Appendix A) L2 (μ) = L2 (μa ) ⊕ L2 (μ s ), where μ is a Borel measure on the circle T, and μa , μ s denote, respectively, the absolutely continuous and singular components of μ with respect to the normalized Lebesgue measure m, m{eit : θ1 ≤ t ≤ θ2 } = (θ2 − θ1 )/2π ≤ 1. Lemma 1.3.1 Let T : H → H be a bounded linear operator on a Hilbert space H and let E ⊂ H be a closed subspace. (1) E ∈ Lat(T ) ⇔ E ⊥ ∈ Lat(T ∗ ). (2) E ∈ Lat(T, T ∗ ) ⇔ E ∈ Lat(T ), E ⊥ ∈ Lat(T ). (3) For every E ∈ Lat(T ), E = ER ⊕ E N ,

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.3 Non-reducing Subspaces of the Bilateral Shift Mz

9

where ER ∈ Lat(T, T ∗ ) (a reducing subspace of T ) and E N ∈ Lat(T ) is a completely non-reducing subspace, i.e. such that E  ⊂ E N , E  ∈ Lat(T, T ∗ ) ⇒ E  = {0}. This representation is unique. Proof (1) We first show the implication “⇒”. Let y ∈ E ⊥ . Then, (T ∗ y, x) = (y, T x) = 0 for every x ∈ E, and hence T ∗ y ∈ E ⊥ . It ensues that T ∗ E⊥ ⊂ E⊥. The implication “⇐” is immediate since T = (T ∗ )∗ . (2) It is immediate by (1) since T = (T ∗ )∗ . (3) Clearly the “span” (closed linear hull) of a family of reducing subspaces is still in Lat(T, T ∗ ). Set   ER = span E  : E  ⊂ E, E  ∈ Lat(T, T ∗ ) , E N = E  ER . Then E = ER ⊕ E N and ER ∈ Lat(T, T ∗ ). Moreover, E N = E ∩ (ER )⊥ and hence, by (1), E N ∈ Lat(T ). If E  ⊂ E N and E  ∈ Lat(T, T ∗ ), then E  ⊂ ER by the definition of the latter. Thus E  = {0}. The uniqueness is also immediate.  Lemma 1.3.2 Let μ be a finite Borel measure on T, with μ = μa +μ s = w·m+μ s its Radon–Nikodym decomposition (see Appendix A), and let E ⊂ L2 (μ) be a NON-reducing invariant subspace of Mz : L2 (μ) → L2 (μ). Then: (1) There exists a function q ∈ E such that |q|2 w = 1 m-a.e. (2) ER ⊂ L2 (μ s ), where ER is the reducing part of E according to Lemma 1.3.1. Proof (1) Our subspace E satisfies the properties Mz E ⊂ E, Mz E  E; indeed, if we had Mz E = E, then Mz∗ E = Mz Mz E = E, hence E ∈ Lat(Mz , Mz∗ ) which is not the case. Moreover, Mz is an isometric (and even unitary) operator, and thus the image Mz E is closed. Let q ∈ E  Mz E = E ∩ (Mz E)⊥ ,

q = 1.

Since q ∈ E and Mzn q ∈ Mz E for all n ≥ 1, we obtain   n n z qq dμ = zn |q|2 dμ, 0 = (z q, q) = T

T

n ≥ 1.

We conclude, by complex conjugation, that all the Fourier coefficients of the measure |q|2 dμ, except for one, are zero, and hence there exists a 2 dμ)(n) = cm(n) ˆ for all n, n ∈ Z. By the theorem constant c such that (|q| of uniqueness (see Appendix A), |q|2 dμ = m (c = 1 by the normalization

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

The space H 2 (T): An Archetypal Invariant Subspace

10

q = 1). Thus, |q|2 dμa +|q|2 dμ s = m, and by the uniqueness of the Radon– Nikodym decomposition m = |q|2 dμa = |q|2 wm, which is equivalent to |q|2 w = 1 m-a.e. (2) Let f ∈ ER . Given that ER is reducing and Mz∗ = Mz = Mz−1 , we have zn f ∈ ER ⊂ E for all n ∈ Z. Then zn f = z(zn−1 f ) ∈ Mz E, and by the definition of q we obtain  0 = (zn f, q) = zn f q dμ, ∀n ∈ Z. T

Therefore, f q = 0 μ-a.e., hence μa -a.e., and thus (given that m = |q|2 dμa ) f q = 0 m-a.e. However q  0 m-a.e., hence f = 0 m-a.e., which means f ∈ L2 (μ s ). We thus obtain ER ⊂ L2 (μ s ).  Corollary 1.3.3 Every invariant subspace of L2 (μ) contained in L2 (μ s ) is reducing and can be written E = χA L2 (μ s ) with A a Borel set. Indeed, if E were not reducing, it would contain a function q satisfying  |q|  0 m-a.e., which is impossible. 2

Definition 1.3.4 (the space H 2 (T), the generic non-reducing subspace) Let L2 (T) = L2 (T, m) (normalized Lebesgue measure ). The Hardy space H 2 (T) is defined as the following subspace of L2 (T):  H 2 (T) = f ∈ L2 (T) : fˆ(n) = 0 for all integers n < 0 . Reminder The exponentials (zn )n∈Z = (eint )n∈Z form an orthonormal basis of the space L2 (T), and hence every function f ∈ L2 (T) is the sum of its Fourier series

fˆ(n)zn , f = n∈Z

N fˆ(n)zn (for N → norm-L (T) convergent for the symmetric partial sums n=−N  n ∞), or even for “disordered” sums n∈σ(N) fˆ(n)z where σ(N) ⊂ Z, σ(N)  Z for N → ∞:

fˆ(n)zn = 0. lim f − N L2 (T) 2

n∈σ(N)

With this reminder, we can say

 



H 2 (T) = f ∈ L2 (T) : f = fˆ(n)zn = an zn : |an |2 < ∞ . n≥0

n≥0

n≥0

Moreover, the use of properties of orthogonal bases leads to   H 2 (T) = spanL2 (T) zn : n = 0, 1, . . . .

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.3 Non-reducing Subspaces of the Bilateral Shift Mz

11

By the above, clearly Mz H 2 (T) ⊂ H 2 (T) and Mz H 2 (T)  H 2 (T) (hence, H 2 (T) is a non-reducing invariant subspace). Moreover, H 2 (T) is completely non-reducing since for every f ∈ H 2 (T), f  0, there exists a positive integer n such that Mz∗n f = zn f  H 2 (T). The following theorem shows that this is a generic example: any other completely non-reducing subspace coincides with H 2 (T) up to a factor of correction. This was proved in the 1960s by Henry Helson, professor at the University of California (Berkeley).

Henry Helson (1927–2010), one of the primary experts of his generation in harmonic analysis, was a professor at the University of California (Berkeley), 1955–2010. His work on Hardy spaces and “abstract Hardy spaces” (1960–1965, in collaboration with David Lowdenslager, a mathematician from Yale), as well as his perfectly written research monographs (such as Lectures on Invariant Subspaces (Helson, 1964)) profoundly influenced the development of the subject. A rich personality with an extraordinary range of talent (in particular, he was a violinist and cellist at a professional level), he had a truly singular career: as a dedicated Quaker, he turned down a position in California in 1948, because he refused to take a “loyalty oath” (mandatory in California in the era of McCarthyism), and left for Europe where he continued his studies in Poland (with Szpilrajn), then in Sweden (with Beurling) and in France, at Nancy (with Schwartz, Dieudonn´e, Godement, and Grothendieck).

Theorem 1.3.5 (Helson, 1964) Let μ be a finite Borel measure on T, μ = μa + μ s = wm + μ s , and let E ⊂ L2 (μ) be an invariant subspace of Mz . Then: (1) either Mz E = E, and then E = χA L2 (μ) with A a Borel set, (2) or Mz E  E, and then E = χA L2 (μ s ) ⊕ qH 2 (T), where |q|2 w = 1 m-a.e. and A is a Borel set; such a function q is unique, and so is A (meaning q = λq with λ ∈ T and χA = χA μ s -a.e.). Conversely, each A and q satisfying this equation generate a reducing subspace by the formula E = χA L2 (μ), and a non-reducing subspace by E = χA L2 (μ s ) ⊕ qH 2 (T). The latter subspace is completely non-reducing if and only if χA L2 (μ s ) = {0} (⇔ χA = 0 μ s -a.e.). Proof (1) This is Wiener’s Theorem 1.2.1.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

12

The space H 2 (T): An Archetypal Invariant Subspace

(2) Let E ∈ Lat(Mz ), E  Mz E. By Lemma 1.3.2, there exists a function q ∈ E ∩ L2 (μa ), q ⊥ Mz E such that |q|2 w = 1 m-a.e. The sequence (zn q)n∈Z is orthonormal:    n m n−m 2 n−m 2 z |q| dμ = z |q| w dm = zn−m dm = δm,n , (z q, z q) = T

T

T

where δm,n is the Kronecker delta (= 0 if m  n, and = 1 if m = n). Consequently,   n 2  an z q : |an | < ∞ . spanL2 (μ) zn q : n ≥ 0 = n≥0

n≥0

Moreover, clearly the mapping U : f −→ q f is a linear isometry of L2 (T) = L2 (T, m) → L2 (T, μa ):   U f 2L2 (μa ) = | f |2 |q|2 w dm = | f |2 dm =  f 2L2 (T) . T

T

  Hence spanL2 (μ) zn q : n ≥ 0 = U(H 2 (T)) = qH 2 (T) ⊂ E. Let E = E  ⊕ qH 2 (T) where E  = E ∩ (qH 2 (T))⊥ (orthogonal complement in L2 (μ)). For an arbitrary function f ∈ E  ⊂ E, we have   zn f q dμ = (q, zn f ) = 0 for n ≥ 1, and f zn q dμ = ( f, qzn ) = 0 for n ≥ 0, T

T

so f q dμ = 0, and hence f q = 0 μ-a.e. However q  0 μa -a.e., thus f = 0 μa -a.e., and then f ∈ L2 (μ s ). We have shown that E  ⊂ L2 (μ s ), and – because the converse E ∩ L2 (μ s ) ⊂ E  is clear – E  = E ∩ L2 (μ s ). As both E and L2 (μ s ) are Mz -invariant, then Corollary 1.3.3 leads to E  = χA L2 (μ s ). For the uniqueness, let χA L2 (μ s ) ⊕ qH 2 (T) = χA L2 (μ s ) ⊕ q H 2 (T) where 2 |q | w = 1 m-a.e. Then clearly χA L2 (μ s ) = χA L2 (μ s ) and qH 2 (T) = q H 2 (T), hence χA = χA μ s -a.e. The second equation implies q/q ∈ H 2 (T) and q /q ∈ H 2 (T), and since |q| = |q | m-a.e., all the Fourier coefficients of q/q , with the exception of (q/q )ˆ(0), are zero. Hence q/q = constant = λ; clearly |λ| = 1. The rest of the statement is also evident.  Corollary 1.3.6 The space L2 (T, μ) contains a non-reducing invariant subspace E (i.e. Mz E ⊂ E, Mz E  E) if and only if m  μ (i.e. w > 0 m-a.e. on T). Indeed, according to Theorem 1.3.5(2), it is a question of the existence of a function q such that |q|2 w = 1 m-a.e., which is equivalent to the condition of the corollary. 

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.3 Non-reducing Subspaces of the Bilateral Shift Mz

13

1.3.1 H p(T) Spaces Let 1 ≤ p ≤ ∞. The Hardy space H p (T) is defined similarly to the space H 2 (T)  H p (T) = f ∈ L p (T) : fˆ(n) = 0 for any integer n < 0 . The H p (T) spaces share many of the properties of the space H 2 (T), but of course not all of them, and always after some modifications. For example, the exponentials (zn )n∈Z no longer form an unconditional basis in L p (T), p  2, but form a Schauder basis for 1 < p < ∞, as will be seen in Chapter 2, Exercise 2.8.4(f)). Here, we limit ourselves to a short list of initial properties. For more information, see Chapter 2, in particular Exercise 2.8.1 (concerning an analog in L p (T) of the theorems of Beurling and Helson). (1) H p (T) is a closed vector subspace of L p (T). (2) If f, f ∈ H p (T), then f = constant. Indeed, all the Fourier coefficients of f are zero, except perhaps fˆ(0).



(3) If f ∈ H p (T) and f = 0 on a set A ⊂ T, with m(A) > 0, then f = 0. For a proof, see Corollary 2.3.3 below. (4) Let  H−p (T) = f ∈ L p (T) : fˆ(n) = 0 for every integer n ≥ 0 .   Then, H p (T) ∩ H−p (T) = {0} and, if p < ∞, closL p H p (T) + H−p (T) = L p (T). (For p = ∞ see Exercise 2.8.4(i). In fact, for 1 < p < ∞, the sum is already closed, by Marcel Riesz’s Theorem 2.8.4(e)). Indeed, the first equation holds for the reason used for (2), the second  because H p (T) + H−p (T) ⊃ P. (5) The invariant subspaces of L p (T) were described by Srinivasan (1963): let E be a subspace invariant under the shift operator Mz : L p (T) → L p (T). Then: (a) either zE = E, and then E = χA L p (T) for some Borel set A, (b) or zE  E, and then there exists q, measurable on T, with |q| = 1 m-a.e., such that E = qH p (T). The parameters A and q are uniquely defined by E in the same sense as in Theorem 1.3.5.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

14

The space H 2 (T): An Archetypal Invariant Subspace

For the proof, see Exercise 2.8.1 where this analog of Theorem 1.3.5 is a corollary of a more general proposition.

1.4 Beurling “Inner Functions” The special case μ = m is particularly important. Corollary 1.4.1 (invariant subspaces of L2 (T)) Let E be a subspace invariant under the shift operator Mz : L2 (T) → L2 (T). Then: (1) either zE = E, and then E = χA L2 (T) for some Borel set A, (2) or zE  E, and then there exists a function q ∈ L∞ (T), with |q| = 1 m-a.e., such that E = qH 2 (T). The parameters A and q are uniquely defined by E in the same sense as in Theorem 1.3.5. Indeed, this is Theorem 1.3.5 with μ s = 0 and w = 1.



The following even more specialized case, called “Beurling’s theorem,” is important for not only its consequences, but also for its role in the development of the theory of Hardy spaces. Even though the proof given below is totally different from the original proof, we still need the following definition introduced by Beurling: that of a special class of “inner functions” in H 2 (T) which, today, plays a fundamental role in the entire theory. See also the historical remarks in the biographical sketch below, and in §§ 1.9, 2.9, 3.6.

Arne Beurling (1905–1986), was a Swedish mathematician, the author of numerous remarkable works in mathematical analysis and cryptography, and a professor (1937–1954) at the University of Uppsala (a university founded in 1477, where Carl Linnaeus and Anders Celsius worked), and later at the Institute for Advanced Study at Princeton, USA. Simultaneously with Gelfand, he discovered the fundamental principles of Banach algebras and introduced an important class of weighted algebras (“Beurling algebras”), described the invariant subspaces of the isometric shift operator, and (with Malliavin) resolved the problem of the completeness radius of families of exponentials by proving the “multiplier theorem,” important for the uncertainty principle in harmonic analysis. His doctoral students include Carleson, Domar, Esseen, Hall, and Nyman.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.4 Beurling “Inner Functions”

15

Beurling is also famous for having single-handedly (in 1940) deciphered the German Nazi secret code known as Geheimfernschreiber (“secret teleprinter”), based on a machine that could create 1018 different combinations (many more than the Enigma machine, famous for its role in Operation Overlord!). This feat allowed the Swedish secret service to systematically decipher coded messages that were passing through Sweden via a cable linking Norway with Nazi Germany. The invasion plan Barbarossa and the date at which it was to start (June 22, 1941) were intercepted and communicated to the Soviets, but they did not believe the information as its source was not revealed. Definition 1.4.2 (Beurling inner functions) A function on the circle T is said to be inner (in the sense of Beurling) if ϕ ∈ H 2 (T) and |ϕ| = 1 m-a.e. Corollary 1.4.3 (Beurling’s Theorem, 1949) Let E ⊂ H 2 (T) ⊂ L2 (T) be a subspace of H 2 (T), E  {0}. Then, E is Mz -invariant if and only if there exists an inner function q such that E = qH 2 (T). There is a bijective correspondence between Lat(Mz |H 2 (T)) and the set of inner functions q whose first non-zero Fourier coefficient is positive. Indeed, by applying Corollary 1.4.1, we see that case (1) is impossible: if zE = E, then we would have zn f ∈ E ⊂ H 2 (T) for all n ≥ 0 and every  function f ∈ E; however if f  0, f = k≥n ak zk with an  0, we obtain  zn+1 f = an z + k>n ak zk−n−1 and hence zn+1 f  H 2 (T). Consequently, zE  E. In case (2) of Corollary 1.4.1, we have E = qH 2 (T) ⊂ H 2 (T), thus q ∈ H 2 (T), and the result follows.  Corollary 1.4.4 (boundary uniqueness theorem) If f ∈ H 2 (T) and f = 0 on a set A ⊂ T such that m(A) > 0, then f = 0. Indeed, let E f := spanL2 (T) (zn f : n ≥ 0) = closL2 (T) ( f Pa ), the smallest Mz -invariant subspace containing f , where Pa is the space of analytic polynomials, Pa := P ∩ H 2 (T).

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

16

The space H 2 (T): An Archetypal Invariant Subspace

If we suppose f  0, then by Corollary 1.4.3, E f = qH 2 (T) with an inner function q (hence, |q| = 1 m-a.e.). In particular, q ∈ E f , which is impossible, since for any polynomial p,   q − p f 22 ≥ |q − p f |2 dm = |q|2 dm = m(A) > 0, A

A



thus a contradiction.

Furthermore, with regard to the uniqueness theorem (proved by Frigyes and Marcel Riesz in 1916: see the biographical sketch in § 1.5), we can add that in Chapter 3 a more complete (even definitive) description of the subject will be presented. Finally, note that numerous examples of inner functions are known (see Exercises § 1.8.3), and better still, that there exists an intelligible description of all inner functions. This was well known long before Beurling’s theorem (see §§ 1.9, 2.9 for details).

1.5 H2 (μ) Spaces and the Riesz Brothers’ Theorem We begin with the definition of the space H 2 associated with an arbitrary Borel measure μ on the circle T (in place of the Lebesgue measure m) and the Radon– Nikodym decomposition lemma of invariant subspaces. Definition 1.5.1 Let μ be a finite measure on T. The Hardy space associated with μ is defined by   H 2 (μ) := spanL2 (μ) zn : n ≥ 0 = closL2 (μ) Pa , where Pa = Lin(zn : n ≥ 0) is again the space of analytic polynomials. Clearly, H 2 (m) = H 2 (T). Lemma 1.5.2 Let μ be a (finite) Borel measure on T. Then: (1) For every E ∈ Lat(Mz ), with Mz : L2 (μ) → L2 (μ) = L2 (μa ) ⊕ L2 (μ s ), we have E = Ea ⊕ E s where Ea = E ∩ L2 (μa ),

E s = E ∩ L2 (μ s ).

(2) H 2 (μ) = H 2 (μa ) ⊕ L2 (μ s ). Proof (1) By Helson’s Theorem 1.3.5, either E = χA L2 (μ) = χA L2 (μa ) ⊕ χA L2 (μ s ), or E = χA L2 (μ s ) ⊕ qH 2 (T) and qH 2 (T) ⊂ L2 (μa ), and the result follows.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.5 H 2 (μ) Spaces and the Riesz Brothers’ Theorem

17

(2) Since H 2 (μ) ∈ Lat(Mz ), by (1) we have E := H 2 (μ) = Ea ⊕ E s . Then by Corollary 1.3.3, E s = χA L2 (μ s ) with a Borel set A. As 1 ∈ H 2 (μ), we have χA = 1 μ s -a.e., and hence E s = L2 (μ s ). Moreover, by writing 1 = 1a ⊕ 1 s ∈ Ea ⊕ E s , we obtain 1a ∈ Ea (1a = 1 μa a.e.), and hence H 2 (μa ) ⊂ Ea . However the reverse inclusion is evident, since for every f ∈ Ea and every sequence of polynomials pn ∈ Pa converging to f , we have  f − pn 2L2 (μa ) ≤  f − pn 2L2 (μa ) + pn 2L2 (μs ) =  f − pn 2L2 (μ) → 0, hence f ∈ H 2 (μa ).



Remark 1.5.3 What equality (2) in the lemma means is that there is a simultaneous polynomial approximation: ∀ f ∈ H 2 (w · m), ∀g ∈ L2 (μ s ), there exists a sequence of polynomials (pn ) ⊂ Pa such that, simultaneously, pn → f (in L2 (μa )) and pn → g (in L2 (μ s )). The following theorem (usually called the Riesz Brothers Theorem) is a cornerstone in the construction of Hardy spaces and of harmonic analysis on the circle T (moreover, there exist analogs of this statement for other groups such as R, Tn , Rn ; see § 1.9). A priori, this is somewhat unexpected: certain restrictions on the Fourier spectrum σF (μ) of a complex measure μ, where   ˆ = n ∈ Z : μ(n) ˆ  0 , μ(n) ˆ = zn dμ, σF (μ) = supp(μ) T

imply consequences on the size of the (Borel) support of μ. Theorem 1.5.4 (Riesz and Riesz, 1916) Let μ be a complex Borel measure on T, assumed “analytic,” i.e. its Fourier coefficients of negative index are zero:  μ(−n) ˆ := zn dμ = 0, n ≥ 0. T

Then, μ  m (μ is absolutely continuous with respect to m) and μ = hm with h ∈ H01 , where  H01 := f ∈ L1 (T) : fˆ(k) = 0 for k ≤ 0 . Proof Let |μ| be the variation of the measure μ (see Appendix A). Clearly μ  |μ|; let be the corresponding Radon–Nikodym derivative: μ = |μ|. It is  well known that | | = 1 |μ|-a.e. (see Appendix A). Since T zn dμ = T zn d|μ|, the hypothesis on μ means that ⊥ H 2 (|μ|) in the space L2 (|μ|). However, by Lemma 1.5.2, H 2 (|μ|) = H 2 (|μ|a ) ⊕ L2 (|μ| s ), which implies ⊥ L2 (|μ| s ), and hence = 0 |μ| s -a.e. Since, at the same time, | | = 1 |μ| s -a.e., we obtain

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

18

The space H 2 (T): An Archetypal Invariant Subspace

|μ| s = 0. Then, the measure μ = |μ| = |μ|a is absolutely continuous with respect to m, and hence by Radon–Nikodym there exists h ∈ L1 (m) such that μ = hm. Clearly the hypothesis on μ can be translated to h ∈ H01 . 

The brothers Frigyes (Fr´ed´eric) and Marcell (Marcel) Riesz were two pillars of analysis in the twentieth century. They founded various domains of analysis, thus offering a rare example of familial scientific endeavor at such a high level. The elder, Frigyes Riesz (1880–1956), laid the foundations of functional analysis and operator theory as separate disciplines (1910); he was strongly influenced by the ideas of Fr´echet, Lebesgue, and Hilbert. The representation theorem of linear functionals, as well as the Riesz–Fischer theorem, bear his name. He also founded the J´anos Bolyai Mathematical Institute and the journal Acta Scientiarum Mathematicarum (Szeged), and with his student B´ela Sz.-Nagy co-authored an influential text, Lec¸ons d’analyse fonctionelle. Marcel Riesz (1886–1969) spent (almost) all of his career at the University of Lund (Sweden). His contribution to analysis was enormous: his discoveries include Riesz transformation, the Riesz potential, the Riesz (–Bochner) mean, and the Riesz–Thorin theorem. Curiously, in his search for a permanent position, he was classed in second place twice in a row (for different positions), each time behind Torsten Carleman. His doctoral students included Thorin, Cram´er, Hille, Frostman, and H¨ormander. Frigyes and Marcel Riesz wrote only one article together (Riesz and Riesz, 1916): it contains the Riesz brothers’ theorem, which subsequently became so important for harmonic analysis and its applications.

Furthermore, with regard to Theorem 1.5.4, it can be mentioned that the original proof was much more complicated than that above; however, our proof depends on invariant subspaces and on some of the already developed theory, hence it is indirect. An alternative proof is presented below, which is completely elementary and depends only on the definition of absolutely continuous measures.

1.5.1 Elementary Proof of Theorem 1.5.4 (Øksendal, 1971)

 First note that the hypothesis on μ implies T p dμ = 0 for every p ∈ Pa , and then T f dμ = 0 for every function f defined and holomorphic in a

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.5 H 2 (μ) Spaces and the Riesz Brothers’ Theorem

19

disk (1 + )D = D(0, 1 + ), > 0: indeed such a function f is the  sum of a power series normally convergent on T, f (z) = k≥0 fˆ(k)zk (with radius of convergence ≥ 1 + ), hence the series can be integrated term by term. In particular, for every rational function f = p/q (p, q ∈ Pa ) having poles (zeros of the denominator q) in C \ D, we have  f dμ = 0.

T

By the definition of μ  m, we must show that, for every Borel set A ⊂ T, m(A) = 0 ⇒ μ(A) = 0. In fact, it is sufficient to do this only for a closed A = F, because of the regularity of the variation |μ| (see Appendix A). So, let F ⊂ T, with F = F, m(F) = 0. We are going to construct a sequence of rational functions (hn ) such that (1) |hn (z)| ≤ 2 on T, (2) limn hn (z) = χF (z) for all z ∈ T. Then, by the dominated convergence theorem,  0=

T

 hn dμ →

T

χF dμ = μ(F) (when n → ∞),

thus μ(F) = 0, which will complete the proof. Construction of a sequence (hn ): since m(F) = 0, for every n ≥ 1, there exist disks D(zi , ri ), i = 1, . . . , N, such that zi ∈ F ⊂ T,

F⊂



D(zi , ri ),

and

i

N

i=1

ri
|zi |nri − |z − zi | > nri − ri = (n − 1)ri , hence   1 ri z − zi    z − zi − zi nri  < (n − 1)ri = n − 1 · However, the other factors of fn are bounded above by 1 (see (2)): hence | fn (z)| < 1/(n − 1) for every point of F. (4) For z ∈ T \ F, let d = dist(z, F); we have |z − zi | ≥ d > 0 for every i. Writing ⎧ N   ⎫ N  ⎪  ⎪ ⎪ nri zi ⎪ nri zi 1 ⎨

⎬ = log 1 − 1− = exp ⎪ , ⎪ ⎪ ⎩ ⎭ f (z) z − zi z − zi ⎪ i=1

i=1

we observe that for n > 2/d,    nri zi  ≤ nri ≤ 1 < 1 ,  z − zi  d nd 2 and by using | log(1 − w)| ≤ 2|w| for |w| ≤ 1/2, we obtain    N N N



2n

nri nri zi  2  log 1 −  ≤ 2 ≤ , ri < z − zi  |z − zi | d i=1 dn  i=1 i=1 and hence limn (1/ fn (z)) = 1. In conclusion, the functions hn = 1 − fn satisfy all the required properties, which completes the proof. 

1.6 The Past and the Future: The Prediction Problem The problems of prediction, prognosis, and extrapolation of stochastic (random) processes have played an extraordinary role in the history of Hardy spaces. Definition 1.6.1 A discrete time stationary process (also known as a stationary sequence) is a sequence (xn )n∈Z in a Hilbert space H such that the elements of its correlation matrix {(xn , xk )H } depend only on the difference n − k, i.e. (xn , xk ) = (xn+ j , xk+ j ),

∀n, k, j ∈ Z,

and H = spanH (xn : n ∈ Z). A subspace E− = spanH (xn : n < 0) is said to be the past of the process, and E+ = spanH (xn : n ≥ 0) the future of the process. The process is said to

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.6 The Past and the Future: The Prediction Problem

21

be singular (or deterministic) if E− = H, and regular (or non-deterministic) if E−  H. The problem of optimal (quadratic, one step ahead) prediction is to calculate distH (xn , Hn ) = inf xn − x, x∈Hn

where Hn = spanH (xk : k < n) is the past of xn . The main problem concerning random processes is to study the “dependence of the future of a process on its past,” and in particular, to measure the best prediction of its state one or several step(s) ahead. The following theorem introduces the central concept of the theory. Theorem 1.6.2 (Kolmogorov, 1939) Let (xn )n∈Z be a stationary random process, H = spanH (xn : n ∈ Z). Then, there exist a unique Borel measure μ on T and a unitary operator U : H → L2 (μ) such that U xn = zn ,

n ∈ Z.

Conversely, for any μ and every unitary operator U : H → L2 (μ), the sequence (U −1 zn )n∈Z is a stationary process. The measure μ is called the spectral measure of the process. Proof First, observe that for every linear combination of xn , we have



2

2

an xn = an ak (xn , xk ) = an ak (xn+1 , xk+1 ) = an xn+1 . n,k

n,k

This means that the mapping defined by V xn = xn+1 , n ∈ Z,   can be extended by linearity V( an xn ) = an xn+1 to an isometric mapping H → H such that V H is dense in H; hence V H = H and V is unitary. Moreover, xn = V n x0 for every n ∈ Z. By the spectral theorem (Appendix E), there exists a unique Borel measure μ on T and a unitary mapping U : H → L2 (μ) such that U x0 = 1 and V = U −1 Mz U, where Mz is the shift operator on L2 (μ). The rest of the statement is immediate.  Corollary 1.6.3 A stationary process (xn )n∈Z is singular if and only if H−2 (μ) = L2 (μ), with H−2 (μ) = spanL2 (μ) (zn : n < 0) and μ the spectral measure of (xn )n∈Z . The corollary is immediate by the definitions and the theorem.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press



22

The space H 2 (T): An Archetypal Invariant Subspace

Andrey Nikolaevich Kolmogorov (1903–1987) was a Russian mathematician, one of the greatest geniuses in mathematics of the twentieth century, creator of the modern mathematical theory of probability, the KAM (Kolmogorov–Arnold– Moser) theory, the Kolmogorov complexity theory, turbulence theory, etc. Dozens of concepts of mathematics and their applications bear Kolmogorov’s name: the Kolmogorov Aintegral, Kolmogorov’s inequality, the Kolmogorov–Smirnov test in statistics, the Kolmogorov 0–1 law, the Chapman– Kolmogorov equation, the entropy of a dynamic system, etc. Originally a member of Nikolai Luzin’s famous group of students (at the University of Moscow), throughout his career he founded a number of scientific schools in different domains, eventually training a total of 69 doctoral students, of which 18 became members of the Academies of Science of various countries. Among other achievements, Kolmogorov is famous for his solution (with Vladimir Arnold) of Hilbert’s 13th problem (1957). He published more than 300 articles, as well as several books that became classics. He was awarded the Chebyshev Prize (1950), the Balzan Prize (1962), the Wolf Prize (1980), and the Lobachevsky Prize (1986), and was a member of dozens of scientific academies and societies. Luzin wrote to him: Вам дaн высокий дух, и я хочу, чтобы Вы его силы берегли для вещей, которые под силу очень немногим . . . (“You were given a great spirit, and I want you to save its strength to achieve exploits accessible by only a very few”). A caveat from Kolmogorov: “Beware of those said to be ‘good mathematicians’ by engineers and ‘good engineers’ by mathematicians.”

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.6 The Past and the Future: The Prediction Problem

23

Lomonosov Moscow State University.

The study of the problem of prediction starts with a few lemmas, somewhat technical but very useful. This study will be continued in Chapters 2 and 3. In the lemmas, μ always stands for the spectral measure of a stationary process (xn )n∈Z . Lemma 1.6.4 For every n ∈ Z, we have distH (xn , Hn ) = distH (x0 , H0 ) = distL2 (μ) (1, H−2 (μ)) = distL2 (μ) (1, H02 (μ)) := d, where H02 (μ) = : spanL2 (μ) (zn : n > 0) = closL2 (μ) (zPa ). Proof We first use the isometric nature of the operators U and V in the proof of Theorem 1.6.2; then, because p ∈ zPa ⇔ p ∈ Lin(zn : n < 0), for every polynomial p ∈ P, we have 1 − pL2 (μ) = 1 − pL2 (μ) .

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press



The space H 2 (T): An Archetypal Invariant Subspace

24

Lemma 1.6.5 Let μ be a finite Borel measure on T. The following assertions are equivalent. (1) (2) (3) (4) (5) (6) (7)

d = 0 (d is defined in Lemma 1.6.4). 1 ∈ H02 (μ) := spanL2 (μ) (zn : n > 0). 1 ∈ H 2 (μ) := spanL2 (μ) (zn : n < 0). z ∈ H 2 (μ) := spanL2 (μ) (zn : n ≥ 0). H 2 (μ) = L2 (μ). H 2 (μ) = L2 (μ). zH02 (μ) = H02 (μ) and/or zH 2 (μ) = H 2 (μ).

Proof (1) ⇔ (2) since in a metric space X, x0 ∈ clos(A) ⇔ distX (x0 , A) = 0. (2) ⇔ (3) by Lemma 1.6.4. (2) ⇔ (4) since the mapping f −→ z f is unitary on L2 (μ). (4) implies limn z − pn L2 (μ) = 0 for a sequence pn ∈ Pa , and hence limn zq − pn qL2 (μ) = 0 for every q ∈ Pa , thus zPa ⊂ H 2 (μ). Since limn zk − zk−1 pn L2 (μ) = 0 for every k ≥ 1, then by induction zn Pa ⊂ H 2 (μ), for every n ≥ 0. Therefore, P ⊂ H 2 (μ), and we obtain (5): H 2 (μ) = L2 (μ). (5) ⇒ (4) is evident. (3) ⇔ (6) for the same reason that (2) ⇔ (5). Finally, clearly, (5) ⇒ (7); and (7) ⇒ (2), since z ∈ zH02 (μ) implies 1 ∈ H02 (μ)  (and the same manipulation with H 2 (μ)). Lemma 1.6.6 Let μ = wm + μ s be the Radon–Nikodym decomposition of a finite Borel measure on T. Then,  2

d :=

Proof

dist2L2 (μ) (1, H02 (μ))

=

dist2L2 (wm) (1, H02 (wm))

= inf

p∈zPa

T

|1 − p|2 w dm.

By Lemma 1.5.2, H 2 (μ) = H 2 (μa ) ⊕ L2 (μ s ); hence H02 (μ) = zH 2 (μ) = H02 (μa ) ⊕ zL2 (μ s ) = H02 (μa ) ⊕ L2 (μ s ).

Writing 1 = 1a ⊕ 1 s (according to the decomposition L2 (μ) = L2 (μa ) ⊕ L2 (μ s )), we have

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.7 Inner–Outer Factorization and Szeg˝o’s Infimum

25

d2 = dist2L2 (μ) (1a ⊕ 1 s , H02 (μa ) ⊕ L2 (μ s )) = dist2L2 (μa ) (1a , H02 (μa )) + dist2L2 (μs ) (1 s , L2 (μ s )) = dist2L2 (μa ) (1a , H02 (μa )).



The spaces H 2 (μ) are the principal tools used in § 1.7, but the final conclusion on the subject will be made in Chapter 2, § 2.7.2.

1.7 Inner–Outer Factorization and Szeg˝o’s Infimum Recall that in Definition 1.4.2 we defined the inner functions, in the sense of Beurling. We now complete this terminology as follows. Definition 1.7.1 Let f ∈ H 2 (T). It is said to be outer if E f = H 2 (T), where E f := spanH 2 (zn f : n ≥ 0), i.e. E f is the smallest (closed) invariant subspace of Mz containing f . Theorem 1.7.2 (Smirnov, 1928a,b) Every function f ∈ H 2 (T), f  0, can be factorized as f = fin fout , where fin is an inner function and fout is outer. This factorization is unique up  is another inner–outer factorization, then to a constant factor: if f = fin fout   fin = λ fin , fout = λ fout with some λ ∈ T. Proof By Corollary 1.4.3, there exists an inner function q such that E f = qH 2 (T). In particular, f = qg where g ∈ H 2 (T). Let us show that g is outer. Indeed, for every function h ∈ H 2 (T) there exist polynomials pn ∈ Pa such that limn pn f − qh = 0. However, pn f − qh2 = qpn g − qh2 =  |q(pn g − h)|2 dm = pn g − h2 , which shows that h ∈ Eg , and hence T Eg = H 2 (T) (i.e. g is outer). By setting fin = q, fout = g, we obtain the desired factorization. For the uniqueness, suppose there is another factorization: then fin fout =     . Let (pn ) be a sequence of polynomials such fin fout , hence fin f in fout = fout  that limn pn fout − 1 = 0. Since fin f in is a unimodular function, we obtain    pn fout − 1 = pn fout fin f in − fin f in  → 0 as n → ∞. However, pn fout fin f in =   ∈ H 2 (T), and consequently fin f in ∈ H 2 (T). Similarly, fin f in ∈ H 2 (T), pn fout  which gives fin f in = constant (compare with the proof of Theorem 1.3.5), and the result follows. 

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

26

The space H 2 (T): An Archetypal Invariant Subspace

Vladimir Ivanovich Smirnov (1887– 1974) was a Russian mathematician, a representative of the Saint Petersburg school (founded by Chebyshev), and a founder of modern complex analysis at Saint Petersburg. He obtained numerous important results on Hardy spaces (canonical factorization, the Smirnov “class D,” Hardy spaces on Smirnov domains, etc.), as well as in ordinary differential equations and mathematical physics. He is also known for his fivevolume Course of Higher Mathematics, which for years dominated the teaching of mathematics at university level in Russia/USSR. He coauthored works with Friedman, Tamarkin, Lebedev, and others. His notable students include Goluzin, Havin, Kantorovich (Nobel Prize in Economics, 1975), Lozinsky, Sobolev, and Yakubovich. Moreover, Smirnov was renowned for his exceptional personality; he was irreproachable for his nobility, kindness, and generosity, even under the unforgiving circumstances of Russian/Soviet reality in the twentieth century.

The “Twelve Colleges” of the University of Saint Petersburg (the rightmost building).

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.7 Inner–Outer Factorization and Szeg˝o’s Infimum

27

Corollary 1.7.3 (Beurling, 1949) Let f ∈ H 2 (T), f  0. Then, E f = fin H 2 (T). Indeed: E f = closH 2 ( fin fout Pa ) = fin closH 2 ( fout Pa )

(since fin is unimodular)

= fin H (T) 2

( fout is outer).



Theorem 1.7.2 also leads to a crucial development in the problem of L2 optimal prediction (see Theorem 1.7.6 below), i.e. in the expression of the quantity d in Lemma 1.6.6 as a function of the measure μ = wm + μ s (more precisely: of the Radon–Nikodym derivative w = dμ/dm),  |1 − p|2 dμ. d2 = dist2L2 (μ) (1, H02 (μ))2 = inf p∈zPa

T

This last extremal problem appeared in the research of G´abor Szeg˝o in the 1920s, and bears his name: the Szeg˝o infimum. However, we first need a property of outer functions of the type “maximum principle” (for details see Chapter 3 below). Theorem 1.7.4 (Smirnov, 1932) Let f ∈ H 2 (T), f  0. Then the following assertions are equivalent. (1) f is an outer function. (2) ∀g ∈ H 2 (T), g/ f ∈ L2 (T) ⇒ g/ f ∈ H 2 (T). Proof (1) ⇒ (2) Let pn ∈ Pa such that limn pn f − 12 = 0, and suppose that g ∈ H 2 (T) such that g/ f ∈ L2 , i.e. g = f h where h ∈ L2 . Then   |pn g − h| dm = |pn f h − h| dm ≤ h2 pn f − 12 → 0 (for n → ∞). T

T

1

The convergence in L (T) implies the convergence of the Fourier coefficients: ˆ = limn (pn g)ˆ(k). However pn g ∈ H 2 (T) (because H 2 (T) is ∀k ∈ Z we have h(k) ˆ = 0 for every k < 0. Thus h ∈ H 2 (T). Mz -invariant), and hence h(k) (2) ⇒ (1) Let f = fin fout be the inner–outer factorization of f . Then, fout ∈ H 2 (T) and fout / f = f in ∈ L2 (T); hence, by (2), f in ∈ H 2 (T) and of course fin ∈ H 2 (T). As seen several times earlier (for example, in Theorem 1.3.5), this  implies fin = constant: hence f is an outer function. Corollary 1.7.5 (1) If f ∈ H 2 (T) is simultaneously inner and outer, then f = constant.

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

28

The space H 2 (T): An Archetypal Invariant Subspace

(2) If f, g ∈ H 2 (T) are outer and | f | = |g| a.e. on T, then f = λg for some unimodular constant λ. Indeed, for (1), we apply Theorem 1.7.4 to 1 and f , and obtain 1/ f = f ∈ H 2 (T), which, with f ∈ H 2 (T), again implies f = constant. For (2), by setting h = f /g and applying the theorem, we obtain h ∈ H 2 (T), and, by switching the roles of f and g, h ∈ H 2 (T). Hence, h = constant (clearly unimodular). 

G´abor Szeg˝o (1895–1985), a Hungarian–German–American mathematician, is known for his work in classical analysis, such as orthogonal polynomials and Toeplitz operators. After obtaining his doctorate in Budapest under the supervision of Fej´er, he went to Berlin and K¨onigsberg, but then, pressured by the Nazis, he emigrated to the USA. His famous collection of solved problems, with P´olya, Aufgaben und Lehrs¨atze aus der Analysis (1925), served for years as an essential source of training for generations of analysts. He is the author of several other reference monographs. His experiences in Budapest included tutoring the young child prodigy Johannes von Neumann. According to witnesses, Szeg˝o was moved to tears by his first meeting with the young Johannes, so rapid and profoundly complete were the responses of his new student.

Theorem 1.7.6 (Szeg˝o, 1920; Verblunsky, 1936; Kolmogorov, 1941) Let μ = wm + μ s be the Radon–Nikodym decomposition of a finite Borel measure on T. Then: (1) either there does not exist any f ∈ H 2 (T) such that | f |2 = w, and then d = distL2 (μ) (1, H02 (μ)) = 0,

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.8 Exercises

29

(2) or there exists a (unique) outer function F ∈ H 2 (T) such that |F|2 = w, and then ˆ > 0. d = distL2 (μ) (1, H02 (μ)) = |F(0)| Proof Suppose d > 0. Then, distL2 (wm) (1, H02 (wm)) > 0 (Lemma 1.6.6), and hence zH02 (wm)  H02 (wm) (Lemma 1.6.5), which implies that H02 (wm) is an invariant non-reducing subspace of L2 (wm). Helson’s theorem (Theorem 1.3.5) provides a function q such that H02 (μ) = qH 2 (T) and |q|2 w = 1 a.e. on T. In particular, z = q f where f ∈ H 2 (T), which implies | f |2 = |z/q|2 = w. Setting F = fout , we obtain |F|2 = w and   d2 = inf |1 − p|2 |F|2 dm = inf |F − pF|2 dm p∈zPa

T

p∈zPa

T

ˆ 2. = dist2H 2 (F, zH 2 (T)) = PH 2 zH 2 F2 = |F(0)| Conversely, if w = |F|2 with an outer function F ∈ H 2 (T), then the last formula ˆ ˆ shows again that d = |F(0)|. It remains to remark that F(0)  0 for every outer  n ˆ ˆ = function F. Indeed, if we suppose F(0) = 0, we would have F = n≥1 F(n)z  k 2 2 2 2 ˆ z( k≥0 F(k + 1)z ) ∈ zH (T), which implies E F ⊂ zH (T) = H0 (T)  H (T). Thus we obtain a contradiction.  In fact, the last theorem does not resolve the prediction problem: expressing an error d(μ) of the best quadratic prediction of a process as a function of the spectral measure μ. In order to obtain the famous Szeg˝o–Verblunsky– Kolmogorov formula     dμ   log   dm , d(μ) = exp dm T we need to develop a theory of “canonical factorization” of functions H 2 (T). This is the goal of Chapter 2.

1.8 Exercises 1.8.1 The Wold–Kolmogorov Decomposition Let T : H → H be a linear isometry in a Hilbert space H, E ∈ Lat(T ) and W = E  T E. Prove the following. (a) T n W ⊥ T m W for every n  m (n, m ≥ 0) (W is said to be a “wandering subspace”).

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

30

The space H 2 (T): An Archetypal Invariant Subspace

Solution: Let x, y ∈ W and n > m ≥ 0. Then, (T n x, T m y) = (T n−m x, y) = 0 because  T n−m x ∈ T E and y ∈ E.

(b) The subspace E∞ = unitary.

 n≥0

T n E reduces T , and the restriction T |E∞ is

Solution: Let x ∈ E∞ . For every n ≥ 0, there exists xn ∈ E such that x = T n xn , which implies T x = T n+1 xn , and hence, T x ∈ E∞ . Moreover, T n xn = T n+k xn+k ⇒ xn = T k xn+k , which in turn implies xn ∈ E∞ , and in particular, x ∈ T E∞ . Hence, T E∞ = E∞ and T |E∞ is a unitary mapping of E∞ onto itself. For the reduction  property, we have T ∗ x = T ∗ T n+1 xn+1 = T n xn+1 , hence T ∗ x ∈ E∞ .

 (c) The subspace E0 = n≥0 ⊕T n (W) is T -invariant and T |E0 is completely non-unitary (i.e. if E  ⊂ E0 , T E  ⊂ E  and T |E  is unitary, then E  = {0}).    Solution: E0 = {x : x = n≥0 T n wn : wn ∈ W, n≥0 T n wn 2 = n≥0 wn 2 < ∞} (convergence in norm, unique representation, see Appendix C). This implies T E0 ⊂  E0 and n≥0 T n E0 = {0}. If E  ⊂ E0 , T E  ⊂ E  and T |E  is unitary, then E  =   n  n  n≥0 T E ⊂ n≥0 T E 0 = {0}.

(d) The Wold–Kolmogorov decomposition (1939): E = E0 ⊕ E∞ . Solution: Let x ∈ E; then, x ∈ E  E0 ⇔ x ∈ E, x ⊥ T n E  T n+1 E (for every n ≥ 0)  ⇔ (consecutively, with n = 0, 1, . . . ) x ∈ E, x ∈ T E, x ∈ T 2 E, . . . ⇔ x ∈ E∞ .

1.8.2 The Shift Operator M z on L2 (T, μ) Let μ be a finite Borel measure on T and Mz : L2 (T, μ) → L2 (T, μ) the shift operator (translation), Mz f = z f . (a) Let E ∈ Lat(Mz ). Describe its Wold–Kolmogorov decomposition (using Helson’s Theorem 1.3.5). Solution: If E is reducing, Mz E = E, then E∞ = E, W = {0}. Otherwise, by Theorem 1.3.5, E = χA L2 (μ s ) ⊕ qH 2 (T) where |q|2 w = 1 m-a.e. (μ = μ s + wm is the Radon–Nikodym decomposition of μ). Then clearly Mz (χA L2 (μ s )) = χA L2 (μ s ) and W = E  Mz E = qH 2 (T)  qzH 2 (T) = qC (a subspace of dim = 1 containing q). Consequently, E∞ = χA L2 (μ s ) and the completely non-unitary portion of Mz is  Mz |qH 2 (T).

(b) Let μi (i = 1, 2) be finite Borel measures on T. Find a necessary and sufficient condition on μi so that the shift operators S i := Mz : L2 (μi ) → L2 (μi )

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.8 Exercises

31

(i = 1, 2) are unitarily equivalent (i.e. there exists a unitary U : L2 (μ1 ) → L2 (μ2 ) such that US 1 = S 2 U). Solution: Suppose S 1 and S 2 are equivalent and U is a unitary operator such that US 1 = S 2 U. Then, US 1k = S 2k U for every k ∈ Z, and hence, for any polynomial p ∈ P, we have U p = p · U1. By a passage to the limit in the last equation (for the norm L2 on the left, and for the norm L1 on the right) we obtain U f = f U1 for any f ∈ L2 (μ1 ). Then U is unitary, and therefore   | f |2 |U1|2 dμ2 = | f |2 dμ1 , ∀ f ∈ L2 (μ1 ), T

T

which implies that μ1 = |U1| μ2 , hence μ1  μ2 . By swapping the roles of S 1 and S 2 , we obtain μ2  μ1 (thus, the measures are equivalent: μ1 ∼ μ2 ). Conversely, if μ1 ∼ μ2 , then μ1 = hμ2 where h ∈ L1 (μ2 ) and 1/h ∈ L1 (μ1 ) (⇔ h  0 μ2 -a.e.), and the √  mapping U f = f h is unitary: U : L2 (μ1 ) → L2 (μ2 ) and satisfies US 1 = S 2 U. 2

(c) The same question as in (b) but for restrictions S i |H 2 (μi ). Solution: The operators S i |H 2 (μi ) are isometric: they are simultaneously unitary or not, and this is the case if and only if H 2 (μi ) = L2 (μi ). If this last equality holds, the question is already answered in (b); if not, we extend the operator U : H 2 (μ1 ) → H 2 (μ2 ) such that US 1 = S 2 U to a mapping U : L2 (μ1 ) → L2 (μ2 ) with the same relation of commutation by the equation U(zn f ) := zn U f , f ∈ H 2 (μ1 ). The final answer is: S i |H 2 (μi ), i = 1, 2, are unitarily equivalent if and only if μ1 ∼ μ2 and  H 2 (μi ) simultaneously coincide (or not) with L2 (μi ) for i = 1, 2.

(d) Describe the finite Borel measures μ on T for which all the invariant subspaces of the shift operator Mz : L2 (μ) → L2 (μ) are reducing, i.e. Lat(Mz ) ⊂ Lat(Mz∗ ) (⇔ Lat(Mz ) = Lat(Mz∗ )). Solution: By Theorem 1.3.5, there exists a non-reducing invariant subspace if and only if there exists a measurable function q such that |q|2 w = 1 m-a.e. on T, with w = dμ/dm. Clearly the last property is equivalent to w > 0 m-a.e. on T. The answer to (d) is: it is necessary and sufficient that w = 0 on a set σ ⊂ T having m(σ) > 0 (which is equivalent to m  μ). 

1.8.3 Inner and Outer Functions A few “bare-hands” examples, without using the theory of Chapter 2, but nonetheless using knowledge of the multipliers of H 2 (T) (part (a) below).

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

The space H 2 (T): An Archetypal Invariant Subspace

32

(a) Multipliers, algebra H ∞ . Here L2 = L2 (T) and H ∞ (T) := H 2 (T) ∩ L∞ (T). The multiplier spaces are defined by Mult(L2 ) = {h : f ∈ L2 ⇒ h f ∈ L2 }, Mult(H 2 (T)) = {h : f ∈ H 2 (T) ⇒ h f ∈ H 2 (T)}. (i) Show that Mult(L2 ) = L∞ (T), Mult(H 2 (T)) = H ∞ (T). Solution: The inclusion L∞ (T) ⊂ Mult(L2 ) is evident. For the converse, let  h ∈ Mult(L2 ), then ∀ f ∈ L2 T | f |2 |h|2 dm < ∞. However, g = | f |2 is an arbitrary positive function of L1 (T), hence h ∈ L∞ (T) (Appendix A). For the case of Mult(H 2 (T)), clearly Mult(H 2 (T)) ⊂ H 2 (T) and thus, for any h ∈ Mult(H 2 (T)), the multiplication operator Mh f = h f is continuous H 2 (T) → L1 (T), hence it is closed for H 2 (T) → H 2 (T), thus bounded (by the closed graph theorem). Consequently, the formula Mh ( f ) = zn Mh (zn f ) extends Mh on the subspace zn H 2 (T) ⊂ L2 (with the same norm), and by approximation, on the whole space L2 . Thus Mult(H 2 (T)) ⊂ L∞ (T), which leads to Mult(H 2 (T)) ⊂ H ∞ (T). For the converse, note that h ∈ H ∞ (T), p ∈ Pa ⇒ hp ∈ H 2 (T), and again by approximation (letting pn − f 2 → 0), we obtain h f ∈ H 2 (T) for any f ∈ H 2 (T), which shows that H ∞ (T) ⊂ Mult(H 2 (T)). 

(ii) H ∞ (T) is a Banach algebra for standard multiplication on T. Moreover, for every function f ∈ L2 , we have f · H ∞ (T) ⊂ E f . Solution: The space of multipliers Mult(X) = {h : f ∈ X ⇒ h f ∈ X} of a function space X is clearly an algebra. Moreover, the inequality  f g∞ ≤  f ∞ g∞ for f, g ∈ H ∞ is also evident, and the result follows for the algebra H ∞ (T). For the rest, clearly f Pa ⊂ E f . It only remains to show that ( f Pa )⊥ ⊂ ( f H ∞ )⊥ (orthogonal complement  in L2 ). Let g ∈ ( f Pa )⊥ , i.e. T g f p dm = 0 for any polynomial p ∈ Pa . Thus for any  h ∈ H ∞ , T g f h dm = 0 because g f ∈ L1 and h is a weak limit σ(L∞ , L1 ) of its Fej´er polynomials (see Appendix A). 

(b) Examples of inner functions. Show that the following functions are inner. (i) bλ = (λ − z)/(1 − λz) where λ ∈ D = {z ∈ C : |z| < 1}.  n Solution: bλ = (λ − z) n≥0 λ zn (|z| = 1), and clearly bˆ λ (k) = 0 for k < 0, and  2 2 ˆ k≥0 |bλ (k)| < ∞: hence b ∈ H (T). Moreover, for |z| = 1, we have |λ − z| = |λ − z| =  |1 − λz|, thus |bλ (z)| = 1.

(ii) f =

N k=1

bλk where λk ∈ D.

Solution: As H ∞ (T) · H ∞ (T) ⊂ H ∞ (T) (by part (ii) of (a)), a product of inner functions is inner. 

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.8 Exercises

33

(iii) sζ,a = exp(−a((ζ + z)/(ζ − z))) where a > 0, ζ ∈ T. Solution: As

 ζ + z  1 − |z|2 ≥0 Re = ζ−z |ζ − z|2

for any ζ ∈ T, |z| ≤ 1, z  ζ, we obtain |sζ,a | = 1 on T. Moreover, for every n > 0, we   have sˆζ,a (−n) = T zn sζ,a (z) dm = limr→1 T fr (z) dm = 0 where f (z) = zn sζ,a (z) and  fr (z) = f (rz), 0 ≤ r < 1 ( fˆr (0) = 0 since fr is analytic in |z| < 1/r and fr (0) = 0.

(iv) f =

N

k=1 sζk ,ak

where ak > 0, ζk ∈ T.

Solution: See the solution of (ii) above.



(c) Examples of outer functions. Show that the following functions are outer. (i) f ∈ H 2 (T) such that 1/ f ∈ H ∞ (T). Solution: By (a,ii), clearly 1 = f · 1/ f ∈ E f , hence E f = H 2 (T).



(ii) f ∈ H ∞ such that Re( f ) ≥ 0. Solution: For any > 0 there exist (a large) r > 0 and (a small) δ > 0 such that | f + − r| ≤ (1 − δ)r a.e. on T (to verify this, sketch the region in C where the values of f (z) + , |z| = 1 are found), or |( f + )/r − 1| ≤ (1 − δ); this implies the normal convergence of the series

 r f + k = . 1− f +

r k≥0 However, part (ii) of (a) implies (1 − ( f + )/r)k ∈ H ∞ , hence r/( f + ) ∈ H ∞ . By (ii) we have f /( f + ) ∈ E f and even   2 f − 1 dm = 0 lim 

→0 T f +

(by the dominated convergence theorem). Thus 1 ∈ E f , hence f is outer.



(iii) f = 1 + g, g ∈ H ∞ , g∞ ≤ 1. Solution: This is a special case of (ii).



(iv) f ∈ H 2 (T) such that Re( f ) ≥ 0. Solution: The solution of (ii) above shows that it suffices to prove the inclusion 1/( f + ) ∈ H ∞ , or the inclusion 1/( f + ) ∈ H 2 (T) (since 1/( f + ) ∈ L∞ is evident).  To this end, fix 0 < r < 1 and consider fr (z) = k≥0 fˆ(k)rk zk , z ∈ T. Then, fr ∈ C(T)

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

34

The space H 2 (T): An Archetypal Invariant Subspace

(hence bounded) and Re( fr ) ≥ 0 (because fr = f ∗ Pr , which is a convolution with the positive function

1 − r2 Pr (z) = = r|k| zk , 2 |z − r| k∈Z z ∈ T, see Appendix A). By the solution (ii) above, 1/( fr + ) ∈ H ∞ ⊂ H 2 (T) and by the dominated convergence theorem, 1 1 = 0. − lim r→1 fr +

f + 2 

Thus 1/( f + ) ∈ H 2 (T).

(d) An extremal problem. First, we justify Cauchy’s formula for Fourier coefficients: f g(n) = (i) Let f, g ∈ L2 (T) (thus f g ∈ L1 (T)). Show that, for every n ∈ Z,   ˆ g ˆ (k) f (n − k): the series converges absolutely. k∈Z Solution: By Cauchy’s inequality  f (g − g )1 ≤  f 2 g − g 2 , the multiplication  Mg f = f g is continuous L2 (T) → L1 (T). Moreover, the Fourier series g = k∈Z gˆ (k)zk  converges for the norm of L2 (T). Hence, f g = k∈Z gˆ (k)zk f converges in L1 (T),  which implies  f g(n) = k∈Z gˆ (k)(zk f )(n). The calculation (zk f )(n) = fˆ(n − k) is elementary. 

(ii) Let f = fin fout ∈ H 2 (T). Show that sup{|ˆg(0)| : g ∈ H 2 (T), |g| ≤ | f | a.e. on T} = | fˆout (0)|.  ˆ Solution: By (i), clearly ϕ ψ(0) = ϕ(0) ˆ ψ(0) for all functions ϕ, ψ ∈ H 2 (T). ˆ Moreover, for every inner function h, we have |h(0)| ≤ h1 = 1. Given g ∈ H 2 (T), |g| ≤ | f |, which implies |ˆg(0)| = |ˆgin (0)ˆgout (0)| ≤ |ˆgout (0)|. Then by Theorem 1.7.6,   |1 − p|2 |g|2 dm ≤ inf |1 − p|2 | f |2 dm = | fˆout (0)|2 .  |ˆg(0)|2 ≤ |ˆgout (0)|2 = inf p∈zPa

T

p∈zPa

T

1.9 Notes and Remarks As already mentioned, Hardy spaces H p were defined in 1915 (Hardy, 1915), and by 1930 the essentials of the theory had been constructed. At the time, it was a novel mix of fundamental ideas: complex analysis, the Lebesgue integral, and functional vector spaces. Very rapidly, Hardy spaces became one of the mainsprings of the development of analysis in the twentieth century. However, the theory had to wait another 30 years, until the 1960s, for the true

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

1.9 Notes and Remarks

35

magnitude of its potential to be revealed, via the discovery of the main source of its force: the invariant subspaces of the group of translations (Mzn )n∈Z and its semigroup (Mzn )n∈Z+ . Arne Beurling (1949) formulated the correspondence between the invariant subspaces and the inner–outer factorization (in fact, the latter had been known by Smirnov for more than 20 years (Smirnov, 1928a,b)). Beurling’s work led to the discovery of the hidden heart of the theory (Helson and Lowdenslager, 1961; Helson, 1964)): the fact that analyticity is a consequence of the causality of the semigroup under consideration, and that the main feature of the subject is that this semigroup is linearly ordered (it is not very important whether we take Z or R, as made clear in the years 1920–1940). The presentation of this book is based on the novel version of the theory proposed by Helson (1964) (see also Nikolski (1980, 1986, 2002)): in his work, the point of view described above is accepted from the start as the cornerstone of the whole construction. This is a spectacular difference from the classical and/or post-modern presentations, i.e. Privalov (1941), Duren (1970), Garnett (1981), Stein (1993), Koosis (1980), and Pavlovi´c (2004). Formal references: for Theorem 1.2.1 see Wiener (1933), for Theorem 1.3.5 see Helson (1964), and for Corollary 1.4.3 see Beurling (1949). Historically, the astounding success of the approach by invariant subspaces led to the creation of an “abstract complex analysis” where analyticity is defined and studied with the aid of invariant subspaces with respect to a “semigroup” satisfying certain conditions. This theory is well-developed and is highly efficient for the study of functions of several variables, of almost periodic functions, etc.: see Gamelin (1969) and Barbey and K¨onig (1977). The uniqueness theorem Corollary 1.4.4, as well as Theorem 1.5.4, is due to Riesz and Riesz (1916) (for the proof of § 1.5.1 see Øksendal (1971)). Theorem 1.5.4 plays an important role in several applications, in particular for different forms of the uncertainty principle in harmonic analysis. Numerous generalizations and improvements of this theorem are known; for all these subjects, see Havin and J¨oricke (1994). The contents of § 1.6 are taken from Kolmogorov (1941). The inner–outer factorization of § 1.7 was discovered by Smirnov (1928a,b) and published in a minor Russian journal (but in French! See the Russian translation in Smirnov (1988)). There, Smirnov (following Szeg˝o (1921)) speaks of “maximal functions” instead of “outer functions” (he does not introduce a name for the “inner functions”), which finds a strong justification in several forms of the “maximum principle” (Theorem 1.7.4, found in Smirnov (1932), is one of them; for others, see § 3.3–3.4 below). Because of the isolation of Russia after the Bolshevik revolution, followed by Stalin’s Iron Curtain, these

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

36

The space H 2 (T): An Archetypal Invariant Subspace

results remained almost unknown until the 1960s. The other principal result of Beurling (1949) met with the same destiny: Corollary 1.7.3 is an almost immediate consequence of another article by Smirnov (1932). We can also mention that the inner–outer factorization was rediscovered (practically independently of Smirnov or Beurling) by Wiener and Masani in the framework of the theory of linear prediction (by the generalized Wold–Kolmogorov decomposition), under the name of optimal-residual factorization: see Masani (1966). Theorem 1.7.6 (with the formula mentioned at the end of this section) was proved by Szeg˝o (1920) in the case μ = μa , and by Verblunsky (1936) and Kolmogorov (1941) in the general case. The role and significance of Verblunsky’s 1936 paper was overlooked by the community for many decades and was restored by a thorough historical analysis in Barry Simon’s book Orthogonal Polynomials on the Unit Circle, Part 1: Classical Theory (Simon, 2005; see especially pp. 141, 221). Generalizations for continuous-time processes are also due to Kolmogorov, and for vector-valued processes to Kolmogorov, Matveev, and Rozanov (see Rozanov, 1963), as well as Wiener and Masani (1957, 1958). The Wold– Kolmogorov decomposition (Wold, 1938; Kolmogorov, 1941) plays an important role in the analysis of time series (in the prediction of random processes).

https://doi.org/10.1017/9781316882108.003 Published online by Cambridge University Press

2 The H p (D) Classes: Canonical Factorization and First Applications

Topics. Spaces H p (D), Poisson extension, Jensen’s inequality, Fatou’s theorem, the Smirnov canonical factorization, a return to Szeg˝o’s “inf”, weighted approximation, the Hilbert and Hardy inequalities, the harmonic conjugate, the Littlewood subordination principle.

2.1 Fej´er and Poisson Means First, recall the notion of 2π-periodic convolution, convolution on T = R/2πZ (see Appendix A for more details): if μ, ν are two complex Borel measures on T, then μ ∗ ν is the unique complex measure satisfying    f d(μ ∗ ν) = f (st) dμ(s)dν(t) for every function f ∈ C(T). T

T

T

For measures with density μ = f m, ν = gm, where f, g ∈ L1 (T, m), the definition reduces to the convolution of f and g: μ ∗ ν = ( f ∗ g)m, where, for almost all s,  f (st)g(t) dm(t), s ∈ T. f ∗ g(s) = T

For the Fourier coefficients (see Appendix A), μ ∗ ν(n) = μ(n)ˆ ˆ ν(n) for any n ∈ Z. In this chapter, two important approximate identities from harmonic analysis are frequently used (see Appendix A): those of Fej´er and Poisson. Specifically, for k, n ∈ Z+ and 0 < r < 1, we set Dk =

k

j=−k

ei jx =

sin (k + 1/2) x sin(x/2)

(Dirichlet kernel),

37 https://doi.org/10.1017/9781316882108.004 Published online by Cambridge University Press

The H p (D) Classes: Canonical Factorization and First Applications

38

Φn =

n n 

1

| j|  i jx Dk (x) = 1− e n + 1 k=0 n+1 j=−n

n+1 1  sin 2 x 2 (Fej´er kernel), n + 1 sin(x/2)

1 − r2 = r| j| ei jx (Poisson kernel). Pr (x) = P(reix ) := ix 2 |1 − re | j∈Z

=

Notation For f ∈ L1 (T) and 0 < r < 1, let fr = f ∗ Pr . Lemma 2.1.1 Let f, g ∈ L1 (T). Then: (1) f ∗ g = g ∗ f ,  f ∗ g1 ≤  f 1 g1 . (2) If f ∈ L p (T), 1 ≤ p ≤ ∞, then f ∗ g ∈ L p (T) and  f ∗ g p ≤  f  p g1 . (3) If (Eα ) ⊂ L1 (T) is a family satisfying: (i) C := supα Eα 1 < ∞ and (ii) limα Eˆ α (n) = 1 for every n ∈ Z, then lim  f − f ∗ Eα  p = 0 for every function f ∈ L p (T), α

1 ≤ p < ∞ (approximate identity of L p ). (4) If (i) C := supα Eα 1 < ∞, (ii) limα Eˆ α (0) = 1, and (iii) for every δ > 0, limα (supδ≤|x|≤π |Eα (x)|) = 0, then (Ea ) satisfies conditions (i) and (ii) of (3), and hence is an approximate identity. (5) For every n ∈ Z+ and 0 < r < 1, we have f ∗ Dn =

n

fˆ( j)ei jx = sn ( f, x) (a partial sum of f ),

j=−n

f ∗ Φn =

n

j=−n

f ∗ Pr =



n  | j|  i jx 1

fˆ( j) 1 − sk ( f, x), e = n+1 n + 1 k=0

fˆ( j)r| j| ei jx .

j∈Z

(6) (Φn ) and (Pr ) satisfy properties (i) to (iii) of (4) (when, respectively, n → ∞ and r → 1), and hence are approximate identities. Moreover, (Pr )0 0, we Φ have Φn (x) ≤ ((n + 1) sin2 (δ/2))−1 and Pr (x) =

1 − r2 1 − r2 ≤ 2 (1 − r)2 + 4r sin (x/2) 2 sin2 (δ/2)

for 1/2 ≤ r < 1.

The result follows. The following properties are immediate by Lemma 2.1.1.

https://doi.org/10.1017/9781316882108.004 Published online by Cambridge University Press



40

The H p (D) Classes: Canonical Factorization and First Applications

Corollary 2.1.2 (1) For every f ∈ L p (T), 1 ≤ p < ∞, limn  f − f ∗ Φn  p = 0 ( f ∗ Φn are called the Fej´er polynomials of f ) and limr→1  f − fr  p = 0; moreover, 0 < r < ρ < 1 ⇒  fr  p ≤  fρ  p ≤  f  p . (2) If f ∈ L1 (T) and fˆ(n) = 0, ∀n ∈ Z, then f = 0. Notation Let Hol(D) = { f : f is defined and holomorphic in D} the space of holomorphic functions in D; if f ∈ Hol(D) and 0 < r < 1, we set f(r) (z) = f (rz) for |z| < 1/r. Corollary 2.1.3 For every function f ∈ Hol(D), for 1 ≤ p ≤ ∞ and 0 < r < ρ < 1, we have  f(r) L p (T) ≤  f(ρ) L p (T) , and hence the following limit exists (finite or not): limr→1  f(r) L p (T) = sup0 0 and k large enough. Consequently, the family log | f (rk z)| admits an integrable majorant, thus by letting k → ∞, we obtain Jensen’s formula under the hypothesis f ∈ Hol((1 + )D), > 0. If we suppose f ∈ H 1 (and f (0)  0) and δ > 0, we have  

rk log | f (0)| + = log log | f (rk z)| dm ≤ log(| frk | + δ) dm. |λn | T T |λ | 0, f  0, satisfy the Blaschke condition. Proof Replacing if necessary f by f /zn , we can assume that f (0)  0. By Jensen’s formula Lemma 2.3.1, 

r = log log | f (rz)| dm, log | f (0)| + |λn | T |λ | 0). Indeed, clearly for the shift operator Mz on la2 (wn ) we have Mz  ≤ 1 and hence for every ω ∈ H ∞ , ω(Mz ) ≤ ω∞ (by the von Neumann inequality: see Nikolski (2002)). Under the hypothesis ω(0) = 0, this leads to Cω f 2 ≤ | fˆ(0)|2 + ω∞ Cω (S ∗ f )2 ≤ | fˆ(0)|2 + Cω (S ∗ f )2 , which already implies Cω f  ≤  f  (exactly as in the text of § 2.8.6). The Littlewood principle and its generalizations play an important role in the theory of conformal mappings (see Goluzin, 1966), and in the dynamics of composition operators (see Shapiro, 1993).

https://doi.org/10.1017/9781316882108.004 Published online by Cambridge University Press

3 The Smirnov Class D and the Maximum Principle

Topics. Admissible operations on the outer functions, spectrum of an inner function, GCD and LCM of a family of inner functions, analytic extension and spectrum, classes of Nevanlinna and of Smirnov, the conformally invariant framework, the generalized Phragm´en–Lindel¨of principle. This chapter develops the applied potential of the techniques seen in the preceding chapters, in particular the “maximal” property of outer functions (and from there to a very general maximum principle), as well as the rules for finding the “characteristic” function Θ of an invariant subspace E = ΘH p . The latter leads to a study of the arithmetic of the inner functions and of their analytic behavior at the boundary. All these properties will be useful for applications to filtering theory (Chapter 5) and to the study of the distribution of the zeros of the Euler ζ function (Chapter 6).

3.1 Calculus of Outer Functions Recall that in the terminology of § 2.6.9 and Theorem 2.6.1, an outer function f was defined as an element of the space H p , p > 0, satisfying f = λ[h]. In fact, to define a function [h], the integrability of h is not required, but only that of log |h|. Definition 3.1.1 (general outer functions) Let h be a measurable function on T with log |h| ∈ L1 (T). An outer function (of absolute value |h|) is a function f = λ[h] with |λ| = 1 and, as in Theorem 2.6.1,   ζ+z log |h(ζ)| dm(ζ) , z ∈ D. [h](z) = exp T ζ −z 82 https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.1 Calculus of Outer Functions

83

The following list of properties of outer functions is not exhaustive; however, to recapitulate, a few facts already mentioned are nonetheless included.

3.1.1 Properties of Outer Functions (a) An outer function f admits non-tangential boundary limits b f and f ∈ H p (D) ⇔ b f ∈ L p (T). (See Theorem 2.5.1, Corollary 2.6.2.) (b) Let f ∈ H p , p ≥ 1. Then, f is outer if and only if E f = closH p ( f Pa ) = H p (⇔ f is a cyclic function in H p ). (See Exercise 2.8.1(f).) (c) If f ∈ H p and 1/ f ∈ H q , with p > 0 and q > 0, then f is outer. Indeed, by the canonical factorization of Theorem 2.6.5, f = λ1 B1 V1 [ f ] and 1/ f = λ2 B2 V2 [1/ f ], hence 1 = λBV, and by the uniqueness of the factorization  B = B1 B2 = constant, V = V1 V2 = constant, hence f = λ1 [ f ]. (d) Theorem (Smirnov, 1928a,b) (1) Let f ∈ Hol(D) with Re( f ) ≥ 0 in D. Then f is outer and f ∈ H p (D) for every 0 < p < 1 (but perhaps f  H 1 (D). (2) More generally, if f ∈ Hol(D), f (z)  0 and α := supz∈D | arg( f (z))| < ∞, then f is outer and f ∈ H p (D) for every 0 < p < π/2α (but perhaps f  H π/2α (D)).  (3) For every h ∈ L1 (T), Γh ∈ 0 0), but this does not hold for certain other outer functions f ∈ H∞. Indeed, if f ∈ Hol((1 + )D) without zeros in D, then f = pg where p is a polynomial without zeros in D and g±1 ∈ H ∞ (D). Hence, g is outer (see (c)  above) and p = A nk=1 (1 − z/λk ), |λk | ≥ 1, which leads to Re(1 − z/λk ) ≥ 0 for z ∈ D. By (e) and (f), f is outer. If := inf D | f | > 0 and g∞ < , then f + g is outer by part (iii) of (g) above. To justify the last assertion, let f = 1 − z and f = (1 − z)V where  1 + z V = exp −

. 1−z It is easy to see that lim →0  f − f ∞ = 0, that f is outer, but not the  functions f .

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

86

The Smirnov Class D and the Maximum Principle

(j) Here we use certain properties of § 3.3.1, more precisely those of § 3.3.1(d) and § 3.3.1(f); as can be easily assured, the proofs of § 3.3.1(d) and § 3.3.1(f) are independent of § 3.1.1(j). The function e f is outer for any f ∈ H 1 (D). Indeed, when h := Re( f ) ∈ L1 (T), then h(z) = h ∗ Pz and  ζ+z h(ζ) dm(ζ), f = ic + T ζ −z where c ∈ R (see also Exercise 2.8.4(c)). Set hn (ζ) = min(h(ζ), n) and  ζ+z hn (ζ) dm(ζ), Fn = exp( fn ). fn (z) = ic + T ζ −z Then, fn ∈ H p , 0 < p < 1 (see (d) above) and Re( fn (ζ)) = hn (ζ)  h(ζ) = Re( f (ζ)) a.e. on T. Hence, |Fn (z)| = exp(Re( fn (z))) ≤ en and |Fn (z)| = exp(Re( fn (z)))  exp(Re( f (z))) = |e f (z) |. By § 3.3.1(f), we obtain e f ∈ D. With the same argument but for hn (ζ) = max(h(ζ), −n) we obtain e− f ∈ D. Thus e± f ∈ D, which implies that e f is an outer function (see § 3.3.1(d)).  Remark In a way, this result is optimal, as exp(−(1 + z)/(1 − z)) is not outer.

3.2 Calculus of Inner Functions: The Spectrum The first goal of this section is to study the links between the set operations on the Mz -invariant subspaces of H p and the arithmetic of inner functions, having in mind a bijection Θ −→ ΘH p between these two sets. The other is to study the spectrum of the inner part of a function of H p and its links with the analytic properties of this function. For simplicity, we formulate them in the framework of the space H 2 , but they hold without any changes for any arbitrary p. We begin with the arithmetic. Definition 3.2.1 Let Θ, θ be two inner functions and τ a family of inner functions. (1) The function Θ is said to divide θ (notation: Θ | θ) if θ = Θθ where θ is an inner function. (2) The function Θ is said to be the greatest common divisor of τ (notation: Θ = GCD(τ)) if ∀θ ∈ τ, Θ | θ and if for every inner function Θ such that ∀θ ∈ τ, Θ | θ then Θ | Θ.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.2 Calculus of Inner Functions: The Spectrum

87

(3) The function Θ is the least common multiple of τ (notation: Θ = LCM(τ)) if ∀θ ∈ τ, θ | Θ and if for every inner function Θ such that ∀θ ∈ τ, θ | Θ then Θ | Θ . By convention, the constant 1 is an inner function, and LCM(τ) = 0 if there does not exist an inner function divisible by all the inner functions θ ∈ τ.

3.2.1 Properties of the Divisors, GCDs, and LCMs Let Θ1 , Θ2 be two inner functions and τ a family of inner functions. (a) Θ1 | Θ2 ⇔ Θ1 H 2 ⊃ Θ2 H 2 . Proof If Θ2 = Θ1 Θ3 , then Θ2 H 2 = Θ1 Θ3 H 2 ⊂ Θ1 H 2 . Conversely, if Θ2 H 2 ⊂ Θ1 H 2 , then Θ2 ∈ Θ1 H 2 , hence Θ2 = Θ1 f where f ∈ H 2 . Clearly f is inner.   (b) θ∈τ (θH 2 ) = Θ1 H 2 where Θ1 = LCM(τ), and span(θH 2 : θ ∈ τ) = Θ2 H 2 where Θ2 = GCD(τ).  Proof E := θ∈τ (θH 2 ) is an invariant subspace and hence, if E  {0}, there exists an inner function Θ such that E = ΘH 2 . By definition, for every inner function θ ∈ τ, θ | Θ1 so that Θ1 H 2 ⊂ θH 2 , hence Θ1 H 2 ⊂ E. Moreover, for every inner function θ ∈ τ, θH 2 ⊃ ΘH 2 , so that θ | Θ, thus Θ1 | Θ, giving Θ1 H 2 ⊃ ΘH 2 = E. Hence, Θ1 H 2 = E. If E = {0}, the formula remains valid by the convention of Definition 3.2.1. Similar reasoning proves the second formula.  (c) An invariant subspace generated by a family of functions. Let F ⊂ H p , p ≥ 1, and let EF = spanH p (zn f : n ≥ 0, f ∈ F ) be the subspace of H p generated by F . Then, E = ΘH p where Θ = GCD( fin : f ∈ F ). Proof For every function f from (b).

∈ F , Ef

=

fin H p , and the rest follows 

To find the explicit expressions of the GCD and LCM of a family of inner functions, we use the notation introduced in Remark 2.4.4 and Corollary 2.6.4 for an inner function Θ:      ζ+z dμ(ζ) , z ∈ D, bλ (z)k(λ) exp − Θ(z) = Bk (z)Vμ (z) = T ζ −z λ∈D  where k = kΘ is a zero divisor satisfying the Blaschke condition λ∈D kΘ (λ)(1− |λ|) < ∞ and μ = μΘ a measure on T singular with respect to m. In property (d)

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

The Smirnov Class D and the Maximum Principle

88

below, we also permit divisors k1 that do not satisfy the Blaschke condition; in this case, we set Bk1 = 0. Similarly for the measures: if μ1 (T) = ∞, we set Vμ1 = 0. (d) For any divisors k1 , k2 , we have Bk1 Bk2 = Bk1 +k2 , and for any singular measures μ1 , μ2 , we have Vμ1 Vμ2 = Vμ1 +μ2 . Moreover, if Θ1 = Bk1 Vμ1 and Θ2 = Bk2 Vμ2 , then Θ1 | Θ2 if and only if Bk1 | Bk2 and Vμ1 | Vμ2 , and if and only if k1 ≤ k2 and μ1 ≤ μ2 (meaning μ2 − μ1 ≥ 0). 

Proof Clear. (e) Let Θ1 = LCM(τ) and Θ2 = GCD(τ). Then, Θ1 = Bk1 Vμ1 and Θ2 = Bk2 Vμ2 , where k1 (λ) = sup kθ (λ), θ∈τ

μθ (Aθ ), μ1 (A) = sup

k2 (λ) = inf kθ (λ) (λ ∈ D), θ∈τ

μ1 (A) = inf μθ (Aθ ),

θ∈τ

θ∈τ

where sup and inf are taken over all finite Borel partitions of A (A an arbitrary Borel subset of T), i.e. ∪θ∈τ Aθ = A, Aθ ∩ Aθ = ∅ for θ  θ and Aθ = ∅ for all but a finite number θ ∈ τ. Proof The expressions for k1 and k2 are clear by the division property (d), as are the formulas for μ1 and μ2 , since (again by (d)) they correspond, respectively, to sup (the upper bound) and inf (the lower bound) in the set of positive measures ordered by the relation μ ≤ ν (⇔ ν − μ ≥ 0); see Appendix A.  In the the rest of this section, we study the relationship between the holomorphic extension of a function of H p and the “size” of its inner factor. We know from complex analysis that an analytic function is “well-defined” by its behavior on neighborhoods of the points where it loses analyticity, hence at its singularities. By applying this maxim to 1/ f where f is an inner function, we obtain the notion of the spectrum, i.e. the set of singularities, of an inner function. Definition 3.2.2 Let Θ = BV be an inner function with its canonical factorization, B = Bk (k its zero divisor: see Remark 2.4.4) and V = Vμ (μ is a singular measure on T: see Corollary 2.6.4). The spectrum of Θ is defined by σ(Θ) = supp(k) ∪ supp(μ), where supp(k) = clos{z ∈ D : k(z) > 0} (closure of the zero set of Θ).

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.2 Calculus of Inner Functions: The Spectrum

89

Note immediately that σ(Θ) = ∅ ⇔ Θ = constant, and that σ(Θ1 Θ2 ) = σ(Θ1 )∪ σ(Θ2 ) where the Θ j are inner functions. Both are immediate consequences of the uniqueness of the factorization of Θ j : see Remark 2.4.4 and Corollary 2.6.4. Theorem 3.2.3 (spectrum of an inner function) Let Θ be an inner function and ζ ∈ T. The following assertions are equivalent. (1) ζ  σ(Θ). (2) Θ admits an analytic extension in a neighborhood of ζ. (3) There exists a neighborhood Uζ of ζ such that

  D > 0.

:= inf |Θ(z)| : z ∈ Uζ Proof (1) ⇒ (2) Let Θ = Bk Vμ be the canonical factorization of Θ. Then have σ(Θ) = σ(Bk ) ∪ σ(Vμ ), hence ζ ∈ T \ σ(Bk ) and thus B can be analytically extended in a neighborhood of ζ (since Bk exists and is holomorphic in C \ clos{1/λ : k(λ) > 0}, see Theorem 2.4.2). The same is evident for Vμ because ζ  σ(Vμ ) = supp(μ) and    ζ+z dμ(ζ) . Vμ (z) = exp − supp(μ) ζ − z This proves property (2). (2) ⇒ (3) Clear: as a consequence of (2), Θ is continuous on a neighborhood Uζ ∩ D of ζ and |Θ| = 1 on Uζ ∩ T. (3) ⇒ (1) It suffices to show that ζ  σ(Bk ) and ζ  σ(Vμ ). Given the hypothesis, the first non-inclusion is evident. For the second, suppose that Uζ is an open set satisfying (3) and Δ ⊂ T a closed arc such that Δ ⊂ Uζ (see diagram overleaf). Set    ζ+z dμ(ζ) , z ∈ D. Vμ|Δ (z) = exp − Δ ζ −z Then, by the hypothesis, |Vμ|Δ (z)| ≥ |Vμ (z)| ≥ |Θ(z)| ≥ for z ∈ Uζ ∩ D (). Also inf z∈D\Uζ |Vμ|Δ (z)| > 0 (since dist(D \ Uζ , Δ) = inf{|ζ − z| : ζ ∈ Δ, z ∈ Uζ ∩ D} > 0), and hence 1/Vμ|Δ ∈ H ∞ , which implies Vμ|Δ = constant. By the uniqueness of the Herglotz representation (see Theorem 2.6.3 and  Corollary 2.6.4), we obtain μ | Δ = 0, thus ζ  σ(Vμ ). Corollary 3.2.4 Let Θ be an inner function. Then #  " clos(z : |Θ(z)| ≤ ). σ(Θ) = ζ ∈ D : lim inf |Θ(z)| = 0 =

→0 |z|0

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

90

The Smirnov Class D and the Maximum Principle

Disk D(0,1)

The arc Δ in the set Uζ , where Θ is separated from zero.

Theorem 3.2.5 Let f ∈ H p ( f  0), p > 0, and f = fin fout its inner–outer factorization. Then, f admits an analytic extension at a point ζ ∈ T if and only if so do the functions fin , fout . Proof The sufficiency is evident, so we turn to the necessity. Suppose that f is analytically extendable at a neighborhood of a point ζ ∈ T. It suffices to show that fin is analytically extendable at a neighborhood of ζ. Use the same letter f to denote the analytic extension of f in an open neighborhood Uζ of ζ and by k the multiplicity of the zero of f at ζ. Then, f (z) = (z − ζ)k g(z) where g is holomorphic in D and in the same neighborhood of ζ, and g(ζ)  0. By Theorem 2.6.7(f), g ∈ H p and since (z − ζ)k is outer (see § 3.1.1(i)), we obtain  fin = gin . Consider a sequence of open neighborhoods Uζ ⊂ U ζ ⊂ Uζ ⊂ 



U ζ ⊂ Uζ . Then the function g is bounded and separated from zero on U ζ , say    0 < ≤ |g| ≤ C on T U ζ , and hence for z ∈ D Uζ ,  log |gout (z)| =

 T

Pz (t) log |g(t)| dm(t) =

≤ log C







Pz (t) log |g(t)| dm(t)

T∩U ζ

 

T∩U ζ

Pz (t) dm(t) +



≤ log C +

 +



T\U ζ

T\U ζ



T\U ζ

Pz (t) log |g(t)| dm(t) ≤ log C + C1 ,

since the last integral is well-defined and continuous on C \ (T \ Uζ ), and in particular, bounded on clos(D ∩ U  ) ⊂ C \ (T \ Uζ ). Conclusion: gout (as well

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.2 Calculus of Inner Functions: The Spectrum as g) is bounded on D



91

Uζ . Consequently,

|gin (z)| =

|g(z)| ≥ for z ∈ D ∩ Uζ , |gout (z)| CeC1

and Theorem 3.2.3 implies that gin = fin can be analytically extended  on Uζ .

3.2.2 Logarithmic Residues Given a function f ∈ H p , p > 0, there exists a simple method to find the discrete part of the singular measure μ fin directly as a function of the values of f , without having to find the canonical factorization of f . In fact, for a Borel measure μ on T, we can detail the Radon–Nikodym decomposition μ = μ s + μa by separating in μ s the point masses and the continuous part: μ s = μd + μ sc ,  where μd = t∈T μ({t})δt (a discrete part of μ s , or indeed of μ; in fact, μ({t}) = 0 for every t ∈ T with the exception of a set at most countable in size) and μ sc is a continuous singular measure, i.e. μ sc ({t}) = 0, ∀t ∈ T. To calculate μ fin ({t}) we use the asymptotic behavior of f (z) as z approaches t. In the following theorem, we first consider the case of an isolated singularity t ∈ T (to be used in the next step), and then the general case, where we rely on a corollary of the generalized maximum principle of § 3.4 presented in Exercise 3.5.3. There is no risk of a vicious circle since this part of Theorem 3.2.6 will never be used before § 3.5.4). Theorem 3.2.6 (logarithmic residue) Let f ∈ H p ( f  0), p > 0, with the canonical factorization f = λBV[ f ], and ζ ∈ T. (1) If ζ  σ(B) then μV ({ζ}) = −(1/2) lim(1 − r) log | f (rζ)|. r→1

(2) In the general case, μV ({ζ}) = −(1/2) lim(1 − r) log | f (rζ)|. r→1

This limit is called the logarithmic residue of f at the point ζ. Proof Without loss of generality we can suppose ζ = 1. Let μ = log | f |·m−μV , a real-valued measure (finite, since log | f | ∈ L1 (T)). Then  1 − r2 dμ(t), (1 − r) log | f (r)| = (1 − r) log |B(r)| + (1 − r) |1 − rt|2 T

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

The Smirnov Class D and the Maximum Principle

92

where (1 − r)

1 − r2 ≤ 1 + r ≤ 2 and |1 − rt|2

lim(1 − r) r→1

1 − r2 = 2χ{1} (t) |1 − rt|2

for every t ∈ T. By the dominated convergence theorem,  lim r→1

T

(1 − r)

1 − r2 dμ(t) = −2μV ({1}). |1 − rt|2

(1) In the case where 1  σ(B), clearly limr→1 log |B(r)| = 0, and hence the formula is established. (2) In the general case, it only remains to show that limr→1 (1 − r) log |B(r)| = 0. Suppose limr→1 (1 − r) log |B(r)| < 0. In this case, there exists α > 0 such that (1 − r) log |B(r)| < −α for 0 < r0 < r < 1, and hence the quotient B/S is bounded on the boundary ∂D+ where  α 1 + z S = exp − · 2 1−z and D+ = {z ∈ D : Im(z) > 0}. By Exercise 3.5.3 (a corollary of the Phragm´en–Lindel¨of principle), B/S is bounded in D+ . The same argument for D− = {z ∈ D : Im(z) < 0} shows that B/S is bounded on the disk D, hence S |B, which is absurd.  Example 3.2.7

For every 0 < α < 1 and A > 0, the function   1 + z α  , f = fα (z) = exp −A 1−z

z ∈ D,

is outer (and f ∈ H ∞ ) because it can be extended analytically on C\[0, ∞), and hence σ( fin ) ⊂ {1} and μ fin = μ fin ({1})δ1 . Moreover, limr→1 (1 − r) log | f (r)| = 0 and Theorem 3.2.6(1) give μ fin = 0, and the result follows.

3.3 The Nevanlinna (N) and Smirnov (D) Classes We introduce here two spaces of holomorphic functions, N and D, respectively bearing the names of Rolf Nevanlinna and Vladimir Smirnov, which provide the most general and natural framework for a theory of boundary behavior: in a way, (N) is “maximal” for the existence of non-tangential boundary limits, and (D) is “maximal” for having a maximum principle.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.3 The Nevanlinna (N) and Smirnov (D) Classes

93

Definition 3.3.1 The Nevanlinna class N and the Smirnov class D = N+ are defined as follows:

  Hp , N = f ∈ Hol(D) : f = f1 / f2 where f1 , f2 ∈

D = f ∈ Hol(D) : f = f1 / f2 where f1 , f2 ∈

p>0



 H p and f2 is outer .

p>0

Rolf Nevanlinna (1895–1980), a Finnish mathematician, was one of the key figures in complex analysis of the twentieth century. After defending his thesis in Helsinki in 1919 (under the supervision of Ernst Lindel¨of, professor at the University and a cousin of his father), he became famous as the author of the value distribution theory of meromorphic functions (1925), culminating in his influential monograph Eindeutige analytische Funktionen (1936). President of the International Mathematical Union from 1959 to 1963, throughout his career he received numerous signs of professional recognition (he was an honorary professor of several universities, member of various academies, etc.). The Nevanlinna/Neovius family produced professional mathematicians over at least five generations. Rolf Nevanlinna’s grandfather was a majorgeneral in the army of the Russian Empire (Finland was a province of Russia until 1917) and a professor of mathematics at the military academy. Rolf’s brother Frithiof was also a renowned mathematician, as were two of his descendants (son and grandson). His father, Otto Neovius, was professor of astronomy at the Pulkovo observatory (Saint Petersburg) and a player in the patriotic movement for the liberation of Finland. In 1906 Neovius changed his family name from Swedish to its Finnish translation (Nevanlinna, which can be translated as “Neva river”). In

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

94

The Smirnov Class D and the Maximum Principle

keeping with the same tradition, Nevanlinna was involved in many social movements during his career, often under the colors of the extreme right: a member of the People’s Patriotic Movement, from 1942 to 1943 he presided over the Committee of Volunteers of the Waffen SS. Finally, in 1946, his Germanophile position cost him his position as President of the University of Helsinki. One of his five children (the renowned architect Arne Nevanlinna, born in 1925) gave an unflattering portrait of his father in his book Isan maa (The Land of the Fathers). Several mathematical objects are named after Nevanlinna: the Nevanlinna value distribution theory, Nevanlinna–Pick interpolation, Nevanlinna meromorphic functions, the Nevanlinna characteristic, etc., and also the Nevanlinna Prize awarded every four years at the International Congress of Mathematicians (since 1982). He supervised 28 doctoral theses; among his students were Ahlfors (Fields Medal 1936, the very first after its creation), Karhunen, and Lehto.

3.3.1 A Few Properties of N and D, by Smirnov (1932) (a) D ⊂ N, and both N and D are sub-algebras of the algebra Hol(D); H p (D) ⊂ D (∀p > 0); the functions of N admit non-tangential boundary limits a.e. on T; if f ∈ N and f  0 then log | f | ∈ L1 (T). (b) N = { f ∈ Hol(D) : f = f1 / f2 where f1 , f2 ∈ H ∞ }, D = { f ∈ Hol(D) : f = f1 / f2 where f1 , f2 ∈ H ∞ and f2 is outer}. Proof Let f ∈ N, f = f1 / f2 where f1 , f2 ∈ H p , p > 0, let f1 = λ1 B1 V1 [ f1 ], f2 = λ2 B2 V2 [ f2 ] be their canonical factorizations, and let B = B1 /B2 . Setting g1 = [min(1, | f1 / f2 |)] and g2 = [min(1, | f2 / f1 |)], we obtain g1 , g2 ∈ H ∞ and f = λBV1 g1 /V2 g2 where λ = λ1 /λ2 , which justifies the first formula, as well  as the second: if f ∈ D, we can suppose that V2 = 1. (c) Every outer function (see Definition 3.1.1) is in D. Moreover, D = { f : f = fin fout , fin and fout inner and outer functions, resp.}. A function f ∈ D is outer if and only if  log | f (0)| =

T

log | f (t)| dm(t).

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.3 The Nevanlinna (N) and Smirnov (D) Classes

95

Proof For the outer functions, we use the same verification as before: if f = [h] and log |h| ∈ L1 (T), then f = g1 /g2 where g1 = [min(1, |h|)] and g2 = [min(1, |1/h|)]. The representation of a function of D as the product of the inner and outer parts follows from this and the second formula in (b) above. The criterion with Jensen’s identity follows from (b) and the canonical factorization (see Theorem 2.6.7 for details).  (d) If f1 , f2 ∈ D and f1 f2 is outer, then so are f1 and f2 . (D cannot be replaced by N: f1 = exp(−(1 + z)/(1 − z)), f2 = 1/ f1 .) In particular, if both f and 1/ f ∈ D, the function f is outer. Proof As functions in D, f1 and f2 can be uniquely written in the form f1 = λ1 B1 V1 [h1 ], f2 = λ2 B2 V2 [h2 ] where log |hk | ∈ L1 (T) (k = 1, 2), hence f1 f2 = λ1 λ2 B1 B2 V1 V2 [h1 h2 ] and B1 B2 V1 V2 = constant. By uniqueness, B1 , B2 , V1 , V2 must be constants.  (e) Let f ∈ Hol(D), g ∈ D and | f | ≤ |g| in D. Then, f ∈ D. Proof Indeed, g = g1 /g2 where gk ∈ H ∞ and g2 is outer. By the hypothesis, | f g2 | ≤ |g1 | in D, hence f g2 ∈ H ∞ and f = f g2 /g2 ∈ D, since g2 is outer.  (f) Let f ∈ Hol(D). Then, f ∈ D if and only if there exists a sequence ( fn ), fn ∈ H ∞ (D) such that f (z) = limn fn (z) and | fn (z)|  | f (z)| for every z ∈ D. Proof If f ∈ D, then f = g/h where g, h ∈ H ∞ and h is outer (see (b) above). By § 3.1.1(h), there exists a sequence (hn ) of outer functions such that inf z∈D |hn (z)| > 0 for every n and h(z) = limn hn (z), |hn (z)|  |h(z)| for every z ∈ D. Then the sequence fn = g/hn verifies all the required properties. Conversely, let fn ∈ H ∞ (D) and f (z) = limn fn (z), | fn (z)|  | f (z)| for every z ∈ D, and let fn = λn Bn Vn [ fn ] be the canonical factorizations. We rewrite the condition | fn (z)| ≤ | fn+1 (z)| in the form    Bn (z)Vn (z)  ≤ |[ f / f ](z)|, z ∈ D, n+1 n  Bn+1 (z)Vn+1 (z)  which shows that (Bn Vn /Bn+1 Vn+1 ) ∈ Hol(D) and, by (e), (Bn Vn /Bn+1 Vn+1 ) ∈ D. By the uniqueness of the canonical factorization, we obtain Bn+1 | Bn and Vn+1 | Vn , hence kn+1 ≤ kn (the corresponding divisors) and μn+1 ≤ μn (the corresponding singular measures). This implies the monotone convergence B(z) := limn Bn (z), |Bn (z)|  |B(z)| and V(z) := limn Vn (z), |Vn (z)|  |V(z)|

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

96

The Smirnov Class D and the Maximum Principle

for every z ∈ D, and similarly for [ fn ]. In particular, | fn |  on T. Let h be the limit and z ∈ D such that B1 (z)  0. Then |B1 (z)V1 (z)[ fn ](z)| ≤ |Bn (z)Vn (z)[ fn ](z)| ≤ | f (z)| < ∞,  hence T Pz (t) log | fn (t)| dm(t) < | f (z)/ B1 (z)V1 (z)| for every n = 1, 2, . . . , which implies log(h) ∈ L1 (T). The properties of the Bn , Vn , and [ fn ] lead to a representation f = λBV[h], thus f ∈ D.  (g) Generalized maximum principle (Smirnov, 1932). Let f ∈ D and let g be an outer function. If | f | ≤ |g| on T, then | f | ≤ |g| in D. In particular, H p (D) = D ∩ L p (T),

0 < p ≤ ∞.

Proof Let f1 , f2 , g1 , g2 ∈ H ∞ where f2 , g1 , g2 are outer and such that f = f1 / f2 , g = g1 /g2 . By the hypothesis, we have | f1 g2 | ≤ |g1 f2 | on T, where g1 f2 is an outer function. Applying part (5) of Theorem 2.6.7, we obtain | f1 g2 | ≤ |g1 f2 | in D, and the result follows. For the formula (“integral maximum principle”), the inclusion H p (D) ⊂  p D L (T) is evident, and the converse follows from what has already been proved by setting g = [ f ]. 

3.4 The Generalized Phragm´en–Lindel¨of Principle In this section we show that the Smirnov maximum principle (§ 3.3.1(g)) contains as special cases a variety of very useful propositions, known collectively as the Phragm´en–Lindel¨of principle. We begin with the definition and a few properties of the spaces N and D in domains of the complex plane C different from D.

3.4.1 The Spaces N and D: Conformally Invariant Versions We study the Jordan domains Ω in the extended complex plane C = C ∪ {∞} defined as the conformal images Ω = ω(D) of the disk D by a bi-holomorphic mapping ω : D → C, continuous up to the boundary and bijective in D = clos D; see Appendix B for some references. In § 3.4, Ω will always denote a Jordan domain.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.4 The Generalized Phragm´en–Lindel¨of Principle

97

(a) Definition H ∞ (Ω) = { f ∈ Hol(Ω) :  f ∞ = sup | f (z)| < ∞}, z∈Ω

N(Ω) = { f ∈ Hol(Ω) : f = g/h; g, h ∈ H ∞ (Ω)}, D(Ω) = { f ∈ Hol(Ω) : f = g/h; g, h ∈ H ∞ (Ω), h is outer}, where g outer in Ω means that g ◦ ω is outer in D. The following properties are immediate by the definitions and § 3.3. (b) Properties (i) N(Ω) = { f : f ◦ ω ∈ N(D)}, D(Ω) = { f : f ◦ ω ∈ D(D)}. (ii) f ∈ D(Ω) ⇔ f (z) = limn fn (z) where fn ∈ H ∞ (Ω), | fn (z)|  | f (z)| (∀z ∈ Ω). (iii) A function f ∈ H ∞ (Ω) is outer if and only if f (z) = limn fn (z) where fn ∈ H ∞ (Ω), inf z∈Ω | fn (z)| > 0 (∀n) and | fn (z)|  | f (z)| (∀z ∈ Ω). (iv) Let f ∈ Hol(Ω) and g ∈ D(Ω), | f | ≤ |g| in Ω. Then f ∈ D(Ω) (evident by § 3.3.1(e)). (v) Let Ω1 ⊂ Ω2 be two Jordan domains and f ∈ Hol(Ω2 ). If f is outer in Ω2 , then f | Ω1 is outer in Ω1 (evident by (iii)). Similarly, f ∈ D(Ω2 ) ⇒ f |Ω1 ∈ D(Ω1 ) (evident by (ii)). (c) Generalized maximum principle Let λ ∈ ∂Ω, f ∈ D(Ω) ∩ C(Ω \ {λ}) and let g ∈ C(Ω \ {λ}) be an outer function in Ω such that | f | ≤ |g| on ∂Ω \ {λ}. Then, | f | ≤ |g| in Ω. Proof Evident by the Smirnov theorem § 3.3.1(g).



3.4.2 Generalized Phragm´en–Lindel¨of Principle In fact, the propositions of Theorem 2.6.1, § 3.3.1(g) and § 3.4.1(c) are already versions of the Phragm´en–Lindel¨of principle, but in applications, the condition f ∈ D(Ω) is often replaced by an upper estimate | f (z)| ≤ M(z) of “outer type,” where M is not necessarily holomorphic but is bounded above by an outer function M(z) ≤ |g(z)|, z ∈ Ω (which already implies f ∈ D(Ω) by part (iv) of § 3.4.1(b)). To formalize this passage we introduce a definition. (a) Definition. Let M and M∗ be two non-negative functions on Ω, and w ∈ C(∂Ω \ {λ}) where λ ∈ ∂Ω, w > 0. The function M∗ is called a Phragm´en– Lindel¨of majorant for a pair M, w if the conditions f ∈ Hol(Ω) ∩ C(Ω \ {λ}), | f | ≤ M in Ω and | f | ≤ w on ∂Ω \ {λ} imply | f | ≤ M∗ in Ω.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

98

The Smirnov Class D and the Maximum Principle

Thus one can say that a Phragm´en–Lindel¨of principle is established for a given pair (M, w), if we can find at least one Phragm´en–Lindel¨of majorant M∗ . The continuity condition C(Ω \ {λ}) is, of course, excessive and simply means that we do not wish to discuss the “boundary values” of the functions in Ω. In applications, often λ = ∞.

Lars Edvard Phragm´en (1863–1937), a Swedish mathematician, is known for important developments, first concerning the classical Liouville theorem (1904), then (with Ernst Lindel¨of, 1908) for the maximum modulus principle, and also for his work for Mittag-Leffler’s journal Acta Mathematica, where he started as a member of staff in 1888. He obtained his doctorate at the University of Uppsala in 1889. He gained a large international recognition by his participation in the discovery of what Jean-Christophe Yoccoz called “a fertile error of Henri Poincar´e” (SMF Gazette, vol. 107 (January 2006)). This concerns an event that took place in 1889, when Poincar´e won a prize awarded by King Oscar II of Sweden and Norway for resolving the question of the stability of the three-body problem in celestial mechanics. The prize was presented, and the article accepted by Acta Mathematica, but the young Lars Phragm´en, tasked with re-reading the proofs, found a certain number of errors (90 pages of remarks and objections for a manuscript of 160 pages!). Poincar´e himself then detected a major gap and withdrew the manuscript. He re-submitted it a year and a half later, significantly enriched and already 270 pages long; it was later to form a major part of his masterpiece M´ethodes nouvelles de la M´ecanique c´eleste, the origin of the theory of dynamical systems, ergodic theory, and chaos theory. Soon after, Phragm´en easily obtained a position as professor at the University of Stockholm, where he stayed until 1903 when he left to work for a private insurance company. Today, Phragm´en is especially known for the Phragm´en–Lindel¨of principle, published in Acta Mathematica in 1908 (a version is presented in §§ 3.4 and 3.5).

Ernst Leonard Lindel¨of (1870–1946) was a Finnish mathematician known for his work in topology and analysis. He obtained his doctorate at the University of Helsinki (Helsingfors at the time, when Finland was controlled by Russia) in 1895 and became professor at the same university

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.4 The Generalized Phragm´en–Lindel¨of Principle

99

in 1902. He is especially known as a topologist (Lindel¨of spaces), but also as a specialist in complex analysis: the Phragm´en–Lindel¨of Principle, published in 1908 in Acta Mathematica, and then the Lindel¨of hypothesis on the Euler ζ function, which has resisted all efforts and remains open. For decades, from 1907 to 1938, he was one of the editors of Acta Mathematica, and was an Honorary Professor of several Scandinavian universities.

(b) Theorem (universal Phragm´en–Lindel¨of principle). Let M and w be two functions as in (a) above and let F ∈ D(Ω), G ∈ N(Ω) ∩ C(Ω \ {λ}) be such that M ≤ |F| in Ω and w ≤ |G| on ∂Ω \ {λ}. Then, either (i) ( f ∈ Hol(Ω) ∩ C(Ω \ {λ}), | f | ≤ M (in Ω), | f | ≤ w (on ∂Ω)) ⇒ f = 0, and then M∗ = 0 is a Phragm´en–Lindel¨of majorant, or (ii) there exists an outer function [w ◦ ω], and then M∗ = [w ◦ ω] ◦ ω−1 is a Phragm´en–Lindel¨of majorant for (M, w). Proof By part (iv) of § 3.4.1(b), for a holomorphic function f , the estimates | f | ≤ M ≤ |F| in Ω imply f ∈ D(Ω). If there exists a function f  0 such that f ∈ Hol(Ω) ∩ C(Ω \ {λ}), | f | ≤ M (in Ω) and | f | ≤ w (on ∂Ω), then | f ◦ ω| ≤ w ◦ ω ≤ |G ◦ ω| a.e. on T. Since f ◦ ω, G ◦ ω ∈ N(D), this implies the existence of an outer function [w ◦ ω]. Applying § 3.4.1(c) to g = [w ◦ ω], we  obtain (ii). If such a function f does not exist, we set M∗ = 0.

3.4.3 Classical Examples In § 3.4.2(b) we spoke of a “universal principle” because all known implementations of the Phragm´en–Lindel¨of principles, beginning with the original theorems of Phragm´en and Lindel¨of (1908), Examples 3.4.3(a)–(b) below and the exercises of § 3.5, are special cases of § 3.4.2(b). Moreover, Theorem 3.4.2(b) explains the true nature of the maximum principles and gives an indication of how to construct them in the domain of interest: these are exactly the majorizations of the Smirnov class D that can be extended from an estimate on the boundary ∂Ω to an inequality over the interior of Ω (using the construction of a Szeg˝o maximal function).

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

100

The Smirnov Class D and the Maximum Principle

(a) Example (Phragm´en and Lindel¨of, 1908) Let Ω = C+ = {z ∈ C : Re(z) > 0} and f ∈ Hol(Ω) be such that | f (z)| ≤ A exp(B|z|α ),

z ∈ Ω,

where A, B > 0 and 0 ≤ α < 1. If f | iR is bounded, then f ∈ H ∞ (Ω) (and hence,  f H ∞ (Ω) =  f L∞ (iR) ). The condition α < 1 is optimal (consider for example f (z) = ez ). Indeed, we apply § 3.4.2(b) with w = G = 1 and F(z) = C exp(Czα ) and C > 0 large enough. For z ∈ Ω, we have |z|α ≤ Re(zα )/ cos(πα/2)) and hence A exp(B|z|α ) ≤ |F(z)| if C is sufficiently grand. Moreover, F is an outer function in Ω, because this is the case for its transplantation to the disk (see Example 3.2.7 above):   1 + ζ α  F ◦ ω(ζ) = exp C , 1−ζ where ω is a conformal mapping of D into C+ (ω(ζ) = (1 + ζ)/(1 − ζ)).



(b) Example (Phragm´en and Lindel¨of, 1908) Let f ∈ Hol(C+ ), C+ = {z ∈ C : Re(z) > 0} and let 0 < α, β < 1 be such that | f (z)| ≤ A exp(B|z|α ), z ∈ C+ , and | f (iy)| ≤ C exp(D|iy|β ), iy ∈ iR = ∂C+ , where A, B, C, D > 0. Then, | f (reit )| ≤ C exp(D rβ cos(βt)),

where D =

π D , r > 0, |t| ≤ . cos(πβ/2) 2

Indeed, we apply § 3.4.2(b) with F(z) = K exp(Kzα ) and K > 0 large enough and G = [w] = C[exp(D|z|β )] = C · exp(D zβ ),

Re(z) > 0.

The fact that F and G are outer has already been mentioned (Example (a)). 

3.5 Exercises 3.5.1 An Improvement of Liouville’s Theorem Let f ∈ Hol(C) and let 0 < α < 2 be such that | f (z)| ≤ A exp(B|z|α ),

z ∈ C,

where A, B > 0, and f is bounded on R and iR.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.5 Exercises

101

(1) Show that f = constant. (2) Show with an example that the condition on the exponent α is optimal. Solution: (1) Indeed, it suffices to verify that the restrictions f |C±± on each quarterplane C±± = {z ∈ C : Re(±z) > 0, Im(±z) > 0} are bounded (and then use Liouville’s theorem: every entire bounded holomorphic function is constant). For this, it suffices to note that the exponential F = exp(Czα ) is an outer function in C±± , and to use the same reasoning as in Examples 3.4.3(a) and 3.4.3(b). This last property is verified in the same manner as in the examples, with the help of a conformal mapping  1 + ζ 1/2 . ω : D → C++ , ω(ζ) = i 1−ζ (2) f (z) = exp(iz2 ).

Joseph Liouville (1809–1882) was a French mathematician, a ´ brilliant student of the Ecole Polytechnique from 1825 to 1827, where he studied under Amp`ere and Arago, and graduated at the age of 18 (!), with Poisson as examiner. After taking a break for health reasons, he taught in different establishments in Paris (for 35–40 hours per week!) before being named ´ professor at the Ecole Polytechnique (1838), and then at the Coll`ege de France (1850) and the Facult´e de Sciences in Paris (1857). While accumulating all these positions, Liouville was also active at the Acad´emie des Sciences (elected in 1838) and at the Bureau des Longitudes (1840). In 1836 he founded the Journal de Math´ematiques Pures et Appliqu´ees (also known as the Journal de Liouville) which played an important role in French mathematical life of the nineteenth and twentieth centuries. Liouville wrote more than 400 articles in analysis, number theory, mathematical physics, and even astronomy: in analysis, the Sturm–

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press



102

The Smirnov Class D and the Maximum Principle

Liouville theory (historically, the foundation of Hilbert’s spectral theory) and a Liouville theorem on entire functions (important but today simple: a generalization is presented in § 3.4); in number theory, an explicit construction of transcendental numbers with the aid of continued fractions (Liouville numbers) and a fundamental theorem on Diophantine approximation; in mathematical physics, the invariance of phase space volume for Hamiltonian dynamics (hence for Newtonian mechanics). A well-known episode in Liouville’s career is linked to the unpublished manuscripts of ´ Evariste Galois (containing a revolutionary idea which led to what is now known as Galois theory): it was Liouville who recovered them after the brutal death of the author, interpreted them, and published them in his journal.

3.5.2 The Case of a Strip (Phragm´en and Lindel¨of, 1908) Let Ω = {z ∈ C : 0 < Im(z) < π} and f ∈ Hol(Ω) be such that | f (x + iy)| ≤ A exp(Beα|x| ),

x + iy ∈ Ω,

where A, B > 0 and 0 < α < 1. Show that if f is bounded on ∂Ω, then f ∈ H ∞ (Ω) and  f H ∞ (Ω) =  f L∞ (∂Ω) . Solution: Indeed, if ω(z) = ez is a conformal isomorphism Ω → C+ = {z ∈ C : Im(z) > 0} and g(ζ) = f ◦ ω−1 (ζ) = f (log ζ), a pullback of f to C+ , then     |g(ζ)| ≤ A exp Beα| log |ζ| | = A exp B max(|ζ|α , 1/|ζ|α ) , ζ ∈ C+ . We conclude the proof in the same manner as in Example 3.4.3(a) with a modification of the function F: we apply § 3.4.2(b) with w = G = 1 and F(ζ) = C exp(C(ζ α + ζ −α )) and C > 0 large enough. Using |ζ α + ζ −α | ≤ 2|ζ|α for |ζ| > 1 and (1 − 2−2α )|ζ α | ≤ |ζ α + ζ −α | for |ζ| < 1/2, we obtain (as in 3.4.3(a)) that for C > 0 large enough:   exp B max(|ζ|α , 1/|ζ|α ) ≤ |F(ζ)|, ζ ∈ C+ . As a product of two outer functions F = C exp(Cζ α ) exp(Cζ −α ) (see Example 3.4.3(a)) F is also outer. 

3.5.3 An Inner Function Which Becomes Outer on a Subdomain Let Ω = D+ = {z ∈ D : Im(z) > 0}

and

  1 + z α  fα (z) = exp A , 1−z

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press

3.5 Exercises

103

A ∈ R, 0 < α < 2. Then, fα is an outer function in Ω (however f1 is inner in D). Solution: Indeed, by choosing a conformal isomorphism ω1 : D+ → C+ = {z ∈ C : Im(z) > 0},

ω1 (z) =

 1 + z 2 1−z

,

α/2 we obtain f ◦ ω−1 ), w ∈ C+ . Next, with ω2 : C+ → D, ω2 (w) = 1 (w) = exp(Aw (w − i)/(w + i) and ω = ω2 ◦ ω1 , we have   1 + ζ α/2  F(ζ) := f ◦ ω−1 (ζ) = exp A i , ζ ∈ D. 1−ζ

We first examine the easy case when 0 < α ≤ 1. In this case,  1 + ζ α/2 ≥0 Re i 1−ζ for every ζ ∈ D, and hence for A < 0 the function F is bounded in D, holomorphic in C \ [1, ∞), and thus σ(Fin ) ⊂ {1}. With the zero logarithmic residue,    1 + r α/2   = 0, lim(1 − r) log |F(r)| = lim(1 − r)A i r→1 r→1 1−r we see that F is outer (and hence, so is f ). If A > 0, we apply the preceding arguments to 1/F. In the case where 1 < α < 2, the function F is no longer bounded, but we can get around this obstacle by, for example, using § 3.1.1(j). Indeed,  1 + z α/2 ∈ H p (D) ϕ := A i 1−z for every p < 2/α where 2/α > 1, hence F = eϕ is outer by § 3.1.1(j).



3.5.4 Division by a Singular Function with a Point Measure Let f ∈ H p (D), p > 0 and  1 + z V = exp −A 1−z where A > 0. Show that V | fin if and only if  A−  | f (r)| ≤ C exp − 1−r for every > 0 and a certain C > 0. Solution: Apply Theorem 3.2.6(2) to see that μ fin ({1}) ≥ A.

https://doi.org/10.1017/9781316882108.005 Published online by Cambridge University Press



104

The Smirnov Class D and the Maximum Principle

3.6 Notes and Remarks As already explained, the principal goal of this chapter is to show the decisive role of the outer functions and the Smirnov class D in maximum-principletype estimations. This role, and the class D itself, as well as its importance for problems of polynomial approximation, were discovered by Smirnov (1932) when studying the applicability of the classical Cauchy and Green formulas in Jordan domains Ω,  f (t) 1 dt, f (z) = 2πi ∂Ω t − z  1 ∂G(t, z) f (z) = |dt|, f (t) 2π ∂Ω ∂n where G is the Green’s function of Ω with a pole at z. The conclusion of Smirnov (1932) is as follows: if the boundary ∂Ω is rectifiable, the formulas are applicable if and only if the integrals converge absolutely and f ∈ D(Ω). Other subjects where the class D plays an important role include the theory of conformal mappings (D intervenes in the definition of Smirnov domains in this theory; see for example Duren (1970) and Goluzin (1966)), the description of ideals in algebras of holomorphic functions, etc. Theorem § 3.1.1(d) and its proof appeared in Smirnov (1928b), the important conformally invariant characterization § 3.1.1(h) is from Smirnov (1932). For the definition of the spectrum of an inner function in Theorem 3.2.3, refer to Nikolski (1986), and for the contents of § 3.2 to Hoffman (1962) or Nikolski (1986). The original definition of the Nevanlinna class N is different from Definition 3.3.1, namely: f ∈ N if and only if f is holomorphic in D and  sup 0 0. n xn  (b) Lemma. With the notation of (a), a sequence X is minimal if and only if, for every k, there exists a functional xk ∈ X ∗ such that xn , xk  = δn,k , where δn,k = 1 if n = k and δn,k = 0 otherwise (this is the Kronecker delta). For every finite linear combination, ,

an xn , xk = ak (an ∈ C), and hence, in the case where X generates X, X = spanX (X) (X is said to be complete in X), the xk are uniquely determined by X. Proof The Hahn–Banach theorem (see Appendix D) implies that for a set A ⊂ X and an element x ∈ X, x ∈ spanX (A) ⇔ ( f ∈ X ∗ , f | A = 0 ⇒ x, f  = 0), and the result follows.  (c) Definition. The functional xk of (b) is called a coordinate functional, and the sequence X is said to be the dual (to X). A pair (X, X ), where X is a minimal sequence and X a dual sequence, is called a biorthogonal pair. (d) Lemma. X is uniformly minimal if and only if there exists a dual sequence such that supn xn  · xn  < ∞. If X is complete in X, then sup xn  · xn  = n

1 . δ(X)

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

110

An Introduction to Weighted Fourier Analysis

Proof By a corollary of the Hahn–Banach theorem, for any x ∈ X and every subspace E ⊂ X, we have distX (x, E) = max{|x, f | : f ∈ X ∗ , f | E = 0,  f  ≤ 1}, hence min{ f  : f ∈ X ∗ , x, f  = 1, f | E = 0} = 1/ distX (x, E). Thus, by choosing a coordinate functional xn of minimal norm, we obtain 1/xn  = dist(xn , spanX (xk : k  n)), or

 dist

 xn 1 , spanX (xk : k  n) = , xn  xn  · xn  

and the result follows.

(e) Definition. Let (X, F ) be a biorthogonal pair in X. With each x ∈ X we associate a generalized Fourier series

x∼ x, xn xn , n

and the partial sums Pk,l x =



x, xn xn .

k≤n≤l

The following least upper bound (if it is finite) is called the basis constant of X (or of F ): b(X) := sup Pk,l . k,l

The quantity m(X) := sup Pk,k  = sup xk  · xk  k

k

is called the uniform minimality constant of X. We have m(X) = 1/δ(X) and m(X) ≤ b(X). Given that Pk,l is a projection on X (i.e. P2k,l = Pk,l ), we have Pk,l  ≥ 1, and if X = H is a Hilbert space the equality Pk,l  = 1 is equivalent to the fact that Pk,l is an orthogonal projection. Hence we always have b(X) ≥ m(X) ≥ 1 and, in a Hilbert space, m(X) = 1 (or b(X) = 1) if and only if X is an orthogonal basis.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.1 Generalized Fourier Series

111

4.1.2 Bases The notion of a basis of a Banach space was introduced in Banach’s founding text of functional analysis, the Th´eorie des op´erations lin´eaires (Banach, 1932).

Stefan Banach (1892–1945), a Polish mathematician, creator of the theory of normed vector spaces, was one of the founders (with Maurice Fr´echet) of twentieth century functional analysis. Son of Stefan Greczek, a simple soldier in the service of the AustroHungarian Empire, who was married to Katarzyna Banach, young Stefan was abandoned by his mother four days after his birth (presumably, as he was never able to be certain on this subject), and then by his father, who entrusted him to some friends. He studied at the Technical University of Lw´ow (1910–1914) (Lw´ow = Lviv = Lvov = Lemberg), and then continued at the Jagiellonian University in Krakow. There, in 1916, in the torment of the First World War, Hugo Steinhaus, who was about to take up a professorship at Lw´ow, was strolling in Krakow’s Planty Park when he heard the words “Lebesgue measure” in a conversation between two youths: at this time and place, almost nobody could be expected to know this combination of words! It was Stefan Banach and his friend Otto Nikodym. This is how the Banach–Steinhaus collaboration began (Banach moved to Lw´ow); it was highly productive and lasted up to the beginning of the Second World War. The mathematical community of Lw´ow was substantial: Steinhaus, Banach, Orlicz, Saks, Mazur, Ulam, Schauder, Mark Kac, and others. They gathered together almost every day at the “Scottish Caf´e” to discuss mathematics, pose problems, and work together. The results of each day were written up in a notebook, the “Scottish Book,” which became famous because of its character of a daily mathematical gazette and its collection

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

112

An Introduction to Weighted Fourier Analysis

of unsolved problems and their associated prizes. Some of these were quite humorous: for example, for the basis problem (find a separable Banach space without a Schauder basis: problem 153 in the book, posed by Mazur in 1936), Per Enflo (who provided an example in 1972) received the prize promised in 1936 – a “live goose.” Banach founded the international journal Studia Mathematica, and then the series Monografie Matematyczne (the first volume was Banach’s celebrated Th´eorie des op´erations lin´eaires, the bible of functional analysis specialists for more than 40 years!). In 1939 he was elected President of the Polish Mathematical Society. After the Nazi occupation of Lw´ow (in June 1941) Banach was arrested, and then released (his thesis advisor, Professor Lomnicki, was killed). During the Nazi occupation (1941–1944), bereft of any teaching activities, Banach gained a living as a lice feeder in the German Institute for Infectious Diseases. When the war was over, Banach resumed his mathematical activities, which he intended to continue at Krakow, but he died from lung cancer in 1945.

Per Enflo receiving a live goose from Stanisław Mazur in 1972 for solving problem 153 (a separable Banach space without a basis) in the “Scottish ´ analysis Book.” The book served as a mathematical diary for the Lwow seminar.

Several mathematical concepts bear Banach’s name: the Banach– Steinhaus, Hahn–Banach, and Banach fixed point theorems, the Banach– Tarski paradox (on the decomposition of a ball), Banach spaces, Banach algebras (also defined and investigated by Israel Gelfand under the name of “normed rings”), the Banach indicatrix, etc.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.1 Generalized Fourier Series

113

(a) Definition. In the case where the set of indices is Z+ , a sequence X is a Schauder basis of X, if for every x ∈ X there exists a unique numerical sequence (an (x)) such that k

lim x − an (x)xn = 0. k

n=0

In the case where the set of indices is Z, a sequence X is a symmetric (respectively, non-symmetric) Schauder basis of X, if for every x ∈ X there exists a unique numerical sequence (an (x)) such that k

lim x − an (x)xn = 0 k n=−k

l  

an (x)xn = 0 . resp. lim x − k,l n=−k

We admit the following classical theorem. (b) Theorem (Banach, 1932). A Schauder basis X of a Banach space X indexed by Z+ is a uniformly minimal sequence and an (x) = x, xn  where xn is a coordinate functional of X, and hence x=



x, xn xn

n≥0

is a Fourier series of x convergent for the norm of X. The same holds for bases indexed by Z, in the sense of non-symmetrical convergence. (c) Lemma. Let (X, F ) be a biorthogonal pair in X. Then: (1) If X is a basis of X, then X is complete in X and F is total on X, i.e. (x ∈ X, x, xn  = 0∀n) ⇒ x = 0. (2) X is a basis of X if and only if X is complete in X and supk≥0 P0,k  < ∞ (when indexed by Z+ ), supk≥0 P−k,k  < ∞ (when indexed by Z and with symmetrical convergence) supk,l≥0 P−k,l  < ∞ (when indexed by Z and with non-symmetrical convergence). Proof Property (2) is immediate by (b) above, the Banach–Steinhaus theorem (see Appendix D), and the fact that x = limk P0,k x for every x ∈ Lin(X) (the linear hull of X), with the obvious modifications when indexed by Z.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

114

An Introduction to Weighted Fourier Analysis

If, for every n, x, xn  = 0, then P0,k x = 0 for any k, hence the property (1). 

4.2 Skew Projections In analysis in a Banach (or Hilbert) space, an important role is played by the metric geometry of the space in question. For a Hilbert space this is a fact we could of course expect (since this is the case for elementary analysis in Rn and Cn , n = 1, 2, 3, . . . , and the relations between two or three elements of a Hilbert space are the same as in Rn and Cn , n = 2, 3), and thus everything is ready to define and exploit geometrical concepts such as angles, orthogonality, etc. In a Banach space, this is not so evident because everything must be derived from metrical relations only. In this short section, we introduce the usage of one of these metrical tools: the skew projections. An angle between subspaces can be defined as a function of the skew projections: see § 4.3 below. Definition 4.2.1 Let L, M be two subspaces of a vector space X such that L ∩ M = {0} (there is no hypothesis of any norm or metric on X). Then, the mapping P = PLM : L + M → X defined by P(x + y) = x

(x ∈ L, y ∈ M)

is called a (skew) projection onto L parallel to M.

4.2.1 Properties of P LM Let L, M be two subspaces of a Banach space X such that L ∩ M = {0}. (a) P = PLM is linear, P | L = id, P | M = 0, P2 = P (these properties justify the name “projection” in Definition 4.2.1). Proof Evident.



(b) PLM is continuous if and only if PLM is continuous (where A = closX (A) is the closure of A ⊂ X). Proof Clear since continuity is equivalent to x ≤ Cx + y for every x ∈ L, y ∈ M.  (c) If L and M are closed subspaces, PLM is continuous if and only if L + M is closed, or L + M = closX (L + M). Proof This is a standard consequence of the closed graph theorem.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press



4.3 The Angle Between the Past and the Future

115

(d) Suppose that X = H is a Hilbert space. Then PLM  = sup x∈L

x x = sup , (I − P M )x x∈L P M⊥ x

where P M is an orthogonal projection on M. Proof PLM  = sup x∈L,y∈M

x x = sup , x + y x∈L inf y∈M x + y

and, clearly, inf y∈M x + y = (I − P M )x.



4.3 The Angle Between the Past and the Future The title of this section refers to one of the main applications of the technique developed in this chapter: stationary processes. In this specific application, the subspaces L and M in the following definition are taken to be the past L = spanH (xn : n < 0) and the future M = spanH (xn : n ≥ 0) of a processes (xn )n∈Z in a Hilbert space H: see § 4.7. Definition 4.3.1 (angle between two subspaces) Let H be a Hilbert space and L, M two subspaces of H. The angle (or, the minimal angle) between L and M is a number A = A(L, M) defined by the properties 0 ≤ A ≤ π/2 and cos(A) = sup x∈L,y∈M

|(x, y)| . x · y

4.3.1 Properties of the Angle Let H be a Hilbert space and L, M two subspaces of H. (a) A(L, M) = A(L, M), and L ⊥ M ⇔ A(L, M) = π/2. 

Proof Clear. (b) cos A(L, M) = P M PL  = PL P M . Proof

If x ∈ L, y ∈ M, then (x, y) = (P M x, y) and hence |(P M x, y)| P x = sup M x · y x x∈L y∈M x∈L P P x P P x = sup M L = sup M L = P M PL . x x x∈L x∈H

cos A(L, M) = sup sup

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press



116

An Introduction to Weighted Fourier Analysis

(c) sin A(L, M) = PLM −1 , where PLM is the skew projection of § 4.2. Proof We have sin2 A(L, M) = 1 − cos2 A(L, M) = 1 − sup x∈L

P M x2 (I − P M )x2 = inf 2 x∈L x x2

1 = = PLM −2 , 2 sup x∈L (x /(I − P M )x2 ) where the last equation follows from § 4.2.1(d).



(d) The skew projection PLM is bounded if and only if P M PL  < 1, and if and only if A(L, M) > 0. 

Proof Apply (c), then (b).

4.4 The Case of the Exponentials: A Reduction to P+ We now turn to the principal subject of this chapter: the exponential bases in the spaces L2 (T, μ), where μ is a finite Borel measure on T. It hence concerns the sequence of exponentials X = E, E = (zk )k∈Z = (eikt )k∈Z , which, for any μ, is complete in L2 (T, μ): L2 (μ) = spanL2 (μ) (zn : n ∈ Z) = closL2 (μ) P. Lemma 4.4.1 (Kolmogorov, 1941) Let μ = μ s + μa = μ s + w · m be a finite Borel measure on T with its Radon–Nikodym decomposition. (1) The family of exponentials E is minimal in L2 (T, μ) if and only if it is minimal in L2 (T, wm), and if and only if 1/w ∈ L1 (T) (in particular, the last point holds if E is a basis in the sense of symmetrical summation). (2) The uniform minimality constant is −1/2 −1/2   1 dm dμ , δ(E) = T T w and the dual sequence is  (zn )  a E = ⊂ L2 (T, μa ) = L2 (T, wm). w

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.4 The Case of the Exponentials: A Reduction to P+

117

(3) E is complete in L2 (T, μ) if and only if μ s = 0. (4) If E is a basis (in the symmetrical sense or not) then μ s = 0. Proof (1) Suppose that E is minimal in L2 (T, μ) = L2 (μ s ) ⊕ L2 (wm). Then, for every n ∈ Z+ , the subspace En := spanL2 (μ) (zk : k > n) is Mz invariant, and moreover En = zn+1 H 2 (μ) ⊃ L2 (μ s ) (for the last inclusion see Lemma 1.5.2(2)). Hence, if xn ∈ L2 (μ) is a dual sequence of E (it is unique given that E is complete in L2 (T, μ)), we have xn ⊥ En , thus xn ⊥ L2 (μ s ) for every n ∈ Z. This implies  zk xn w dm. δn,k = (zk , xn )L2 (μ) = (zk , xn )L2 (wm) = T

xn

xn w

Note that ∈ L (wm) ⊂ L (wm), i.e. ∈ L (T), and thus the function xn w has the same Fourier coefficients as zn , so that xn w = zn and 2

1

1

xn = zn /w. Now the inclusion xn ∈ L2 (wm) gives 1/w ∈ L1 (T), which proves the necessity of the condition. Conversely, if 1/w ∈ L1 (T), then clearly the functions xn = (xn )a + (xn ) s with (xn )a = zn /w and (xn ) s = 0 are biorthogonal in L2 (T, μ) with (zk ), hence E is minimal in L2 (T, μ). (2) The explicit formula above for xn and § 4.1.1(d) show that 1/δ(E) = zn L2 (μ) xn L2 (μ) = 1L2 (μ) 1/wL2 (wm) , which is equivalent to the stated formula. (3) For the completeness of E , we know that, for every n, xn ∈ L2 (wm) and, moreover, if f ∈ L2 (wm) and 0 = ( f, xn )L2 (wm) for every n, then f = 0 as a function of L1 (T) with all of its Fourier coefficients zero. Thus, spanL2 (μ) (xn : n ∈ Z) = L2 (wm), and the result follows. (4) Suppose that E is a basis. We keep the notation of part (1). If x ∈   n∈Z E n , then x ∈ E n , thus (x, xn ) = 0 for every n ∈ Z, and hence    n x = n∈Z (x, xn )z = 0. As was seen in (1), L2 (μ s ) ⊂ n∈Z En , so that  L2 (μ s ) = {0}, i.e. μ s = 0. Corollary 4.4.2 Let a function w ∈ L1 (T) be such that 1/w ∈ L1 (T) and let (xn = zn , xn = zn /w), n ∈ Z be a biorthogonal pair of exponentials in the weighted space L2 (T, wm). If f ∈ L2 (T, wm), then f ∈ L1 (T) and its generalized Fourier series

( f, xn )L2 (w) zn n

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

118

An Introduction to Weighted Fourier Analysis

coincides with the classical Fourier series: ( f, xn )L2 (w) = ! f (n),

∀n ∈ Z. 

Indeed, this is clear by Lemma 4.4.1.

We conclude this § 4.4 by connecting the question of bases of exponentials with the angles and skew projections of § 4.2 and § 4.3. We first introduce notation for “analytic” and “anti-analytic” polynomials: P+ = Pa = Lin(zk : k ≥ 0),

P− = Lin(zk : k < 0),

so that P = P+ + P− . We recall the notation for the partial sum and Riesz  f (n)zn , then projections (see § 4.1.1(e), § 2.8.4(e)): if f ∈ P, f = n∈Z !



! ! Pk,l f = f (n)zn , P+ f = f (n)zn , n≥0

k≤n≤l

where k, l ∈ Z, k ≤ l. Clearly, in the sense of Definition 4.2.1, we have P+ = PP+ P− . Similarly, by defining P− f =

 n 0. (6) AL2 (w) (H+2 (w), H−2 (w)) > 0 where H±2 (w) = closL2 (w) (P± ). (7) P+ is continuous on L2 (wm). Moreover, P+  = (sin(AL2 (w) (P+ , P− )))−1 , P+  ≤ b(E) = sup Pk,l  ≤ min{P+ 2 , 2P+ }. k,l

Proof Clearly (1) ⇒ (2), (2) ⇔ (4) (by the Banach–Steinhaus theorem, see also § 4.1.2(c)), (1) ⇔ (3) and (5) ⇔ (6) ⇔ (7) (by § 4.3.1(d)).

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.5 The Hilbert Operator: The Classical Case of L2 (T)

119

(4) ⇒ (7) since for any f ∈ P there exists k = k( f ) such that P+ f = zk P−k,k z−k f , and hence   P+ f  ≤ P−k,k z−k f  ≤ P−k,k  ·  f  ≤ sup P−k,k   f . k≥0

(7) ⇒ (3) since, for any polynomial f ∈ P, we have Pk,l f = zk P+ z−k f − zl+1 P+ z−l−1 f, and thus Pk,l f  = zk P+ z−k f  + zl+1 P+ z−l−1 f  ≤ 2P+  ·  f , giving b(E) = sup Pk,l  ≤ 2P+ . k,l

The equality for P+  is § 4.3.1(c). The lower estimate of b(E) is evident, since P+ f = liml→∞ P0,l f for every f ∈ P. 

4.5 The Hilbert Operator: The Classical Case of L2 (T) The last prerequisite for the principal result of this chapter (§ 4.6) concerns harmonic conjugation (see § 2.8.4), this time in the Hilbert space L2 (T). Recall that for a real function u ∈ L p (T), 1 < p < ∞, the harmonic conjugation mapping H of § 2.8.4(d) (also called the Hilbert operator) is defined by the properties  = 0. u + iHu ∈ H p (T), Hu(0) Clearly H is linear (over the field R): see § 2.8.4(d), where it was shown that Hu p ≤ A p u p with a certain constant A p < ∞. H can be “complexified” by defining H(u + iv) = Hu + iHv for every function u + iv ∈ L p (T), where u, v are real. Then, H becomes linear over the field C, H : L p (T) → L p (T) and bounded with norm H ≤ 2A p . Lemma 4.5.1 The mapping H of § 2.8.4(d) admits the representation Hf =

1 1 (P+ f − P− f ) − ! f (0), i i

and, in the space L2 (T), satisfies the properties H f 2 ≤  f 2

(∀ f ∈ L2 (T)) and

H f 2 =  f 2

(∀ f ∈ L2 (T), ! f (0) = 0).

Proof Given a polynomial f ∈ P, denote the right side of the formula by f (0) ∈ H p and ! g(0) = 0. Hence, g = H f by g. Then, f + ig = 2P+ f − !

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

120

An Introduction to Weighted Fourier Analysis

the definition of H f cited before the statement of the theorem. Consequently, f (n) = ± ! H f (n)/i for ±n > 0, and we obtain the properties of the norm H f 2  by applying Parseval’s identity H f 22 = n0 | ! f (n)|2 .  Remark 4.5.2 The operator H is often called the Hilbert operator. It can be represented in the form of a singular integral operator by using the formulas of § 2.8.4. Indeed, we have seen in § 2.8.4(d) that for a real function f ∈ L2 (T), H f = Im(Γ f ) where Γ is the Herglotz operator,  ζ+z f (ζ) dm(ζ), z ∈ D. Γ f (z) = T ζ −z Hence, for z = reiθ , 0 < r < 1, we have   2π ζ + z iθ Im Qr (t − θ) f (eit )dt/2π, H f (re ) = f (ζ) dm(ζ) = ζ−z 0 T where

 eit + r  2r sin(t) Qr (t) = Im it . = e −r 1 − 2r cos(t) + r2

Since Γ f ∈ H 2 (T), there exist a.e. on T the boundary limits of H f (reiθ ) as r → 1, and it can be shown that for the right-hand side of the equality, there also exists a limit which is equal to the “Cauchy principal value of the integral”: 

2π

H f (re ) → V.P. iθ

f (eit ) cotan

 t − θ  dt

0

2

2π

.

We will not use this form of the operator H, and refer the reader to the vast theory of singular integrals; see Zygmund (1959) or Duoandikoetxea (2001).

4.6 Exponential Bases in L2 (T, μ) It is easy to show a certain number of measures with density μ = wm, w ∈ L1 (T), such that the exponentials E = (zk )k∈Z form a basis for the space L2 (μ). For example, any w with 0 < inf T w ≤ supT w < ∞ is of this type (see Exercise 4.9.2 for more details). It is much more difficult to find such a weight w with a “singularity” (either inf T w = 0, or supT w = ∞). The very first example was given only in 1950 by Ivan Babenko (see Example 4.6.6 below). The complete resolution of the question came only ten years later with the following result of Helson and Szeg˝o.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.6 Exponential Bases in L2 (T, μ)

121

Theorem 4.6.1 (Helson and Szeg˝o, 1960) Let μ = μ s + w · m be a finite Borel measure on T with its Radon–Nikodym decomposition. The following assertions are equivalent. (1) E = (zk )k∈Z is a Schauder basis of the space L2 (μ) (either symmetrical or non-symmetrical). (2) The Riesz projection P+ is well-defined and bounded on L2 (μ). (3) AL2 (μ) (P+ , P− ) > 0. (4) μ s = 0, w = |h|2 where h ∈ H 2 is an outer function such that distL∞ (T) (h/h, H ∞ ) < 1. (5) μ s = 0 and w = eu+Hv where u, v ∈ L∞ (T) are two real functions, v∞ < π/2. Proof

(1) ⇔ (2) ⇔ (3) by Lemma 4.4.3.

(3) ⇔ (4) If E is a basis, by § 4.1.2(b) it is minimal, and by Lemma 4.4.1 μ s = 0 and 1/w ∈ L1 (T), in particular log(w) ∈ L1 (T). Thus there exists an outer function h ∈ H 2 such that w = |h|2 (see Corollary 2.6.2). Now, we simply calculate cos(AL2 (μ) (P+ , P− )). For any f ∈ P+ , g ∈ P− we have     hh h h f gw dm = f hgh 2 dm = f hgh dm := FG dm. ( f, g)L2 (μ) = h h h T T T T Here, we have F = f h ∈ H 2 , G = gh ∈ H02 = zH 2 (as g ∈ P− = zP+ ). Moreover, since the function h is outer, the set L := {F = f h : f ∈ P+ } is dense in H 2 and the set M := {G = gh : g ∈ P− } is dense in H02 . Letting B2 denote the unit ball of H 2 , B2 = {ϕ ∈ H 2 : ϕ2 < 1}, we obtain that L ∩ B2 is dense in B2 , and M ∩ B2 dense in B20 := zB2 . It is also easy to see that the set of products (L ∩ B2 ) · (M ∩ B2 ) is dense in the unit ball B10 of the space H01 = zH 1 (see § 2.8.2). In view of the isometries  f L2 (μ) = FL2 (m) and gL2 (μ) = GL2 (m) , all this leads to cos AL2 (μ) (P+ , P− ) = sup{|( f, g)L2 (μ) | : f ∈ P+ , g ∈ P− ,  f L2 (μ) < 1,  f L2 (μ) < 1}

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

122

An Introduction to Weighted Fourier Analysis 

   h   = sup  FG dm : F ∈ L ∩ B2 , G ∈ M ∩ B20 h T 

  h  = sup  u dm : u ∈ B10 = Φ | H01 , h T

where the linear functional Φ is defined on the space L1 (T) by the formula  h u dm (u ∈ L1 (T)). Φ(u) = T h 1 ∗ ∞ We now use the duality (L ) = L with respect to the bilinear form ϕ, ψ = ϕψ dm (see Appendix D) and the fact that T

(H01 )⊥ := {ψ ∈ L∞ (T) : ϕ, ψ = 0 ∀ϕ ∈ H01 } = H ∞ (verify!). By the Hahn–Banach theorem (see Appendix D), Φ | H01  = distL∞ (T)

h

 h  , (H01 )⊥ = distL∞ (T) , H ∞ , h h

hence cos(AL2 (μ) (P+ , P− )) = distL∞ (T)

h h

 , H∞ .

Given that (3) ⇔ (1)&(2), it follows that (3) ⇔ (4). (4) ⇒ (5) It follows from (4) that there exists g ∈ H ∞ such that h/h  − g∞ 0, |h/h − g| < 1 − and consequently, |h|2 − gh2  < (1 − )|h|2 a.e. on T. Let ζ ∈ T and a = |h(ζ)|2 , then |a − gh2 | < (1 − )a (outside of a negligible set). If we denote α = arcsin(1 − ) (0 < α < π/2) and A = {z ∈ C : | arg(z)| ≤ α}, then the preceding inequality implies that h2 (ζ)g(ζ) ∈ A (a.e. ζ ∈ T). By Corollary 2.2.5, we can be sure that h2 g(D) ⊂ A, and hence there exists a holomorphic function f = log(h2 g). Set v = − Im( f ). Then |v| = | arg(h2 g)| ≤ arcsin(1 − )
0 and δ, 0 < δ < 1, such that Ω ⊂ D(λ, (1 − δ)λ), and hence 1 Ω ⊂ D(1, 1 − δ). λ This implies λ−1 (h/h)g ∈ D(1, 1 − δ) a.e. on T, and then |λ−1 (h/h)g − 1| < 1 − δ, |λ−1 g − h/h| < 1 − δ a.e. on T. As g ∈ H ∞ , distL∞ (T) (h/h, H ∞ ) ≤ 1 − δ < 1.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press



124

An Introduction to Weighted Fourier Analysis

Definition 4.6.2 A function w ≥ 0 is said to be a Helson–Szeg˝o weight, written w ∈ (HS ), if the equivalent conditions (1)–(5) of Theorem 4.6.1 are satisfied. Corollary 4.6.3 With the notation of Theorem 4.6.1, if w ∈ (HS ) then cos(AL2 (μ) (P+ , P− )) = distL∞ (T) 

P+  = 1 − dist2L∞ (T)

h h

h

, H∞

 , H∞ ,

h −1/2

.

Indeed, the first equality is established in the proof, and the second is a consequence of the first and of Lemma 4.4.3.  Corollary 4.6.4 (continuity of P+  as a function of w = eu+Hv ) Let w = eu+Hv , and let P+ w be the norm of P+ and bw (E) be the the basis constant of E in the space L2 (wm). Then, lim bw (E) = 1,

u∞ →0 v∞ →0

lim P+ w = 1.

u∞ →0 v∞ →0

In particular, if w ∈ (HS ), then lim bw (E) = 1,

→0

lim P+ w = 1.

→0

Indeed, it suffices to follow the last lines of the proof of the implication (5) ⇒ (4) of Theorem 4.6.1. It is clear by the definition of Ω that ∀ > 0, ∃η > 0 such that (u∞ < η, v∞ < η) ⇒ (Ω ⊂ D(1, )). In the proof, is denoted 1 − δ, and hence the very last inequality of the proof is distL∞ (T) (h/h, H ∞ ) ≤ . It follows that lim distL∞ (T) (h/h, H ∞ ) = 0,

u∞ →0 v∞ →0

thus by the second formula of Corollary 4.6.3, lim P+ w = 1.

u∞ →0 v∞ →0

The continuity of bw (E) follows from the last bounds of Lemma 4.4.3.



Remark 4.6.5 It is easy to see that for a non-null projection P : H → H on a Hilbert space (P2 = P) we always have P ≥ 1, and P = 1 if and only if P is an orthogonal projection on a closed subspace of H. Thus, for a sequence E, b(E) = 1 if and only if E is an orthogonal basis. Hence

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.7 Prediction and Hankel Operators

125

Corollary 4.6.4 expresses an intuitively evident fact: the closer w is to 1, the more E resembles an orthogonal basis. Nevertheless, it is instructive to note that in the last corollary, if w is a weight with a singularity (i.e. if at least one of the equations inf T w = 0, supT w = ∞ holds), so is w ( > 0). Example 4.6.6 (a power-like weight: Babenko, 1950) Let wα (ζ) = |1 − ζ|α , 1 α ∈ R. Then, w±1 α ∈ L (T) if and only if |α| < 1. Setting u := log(wα (ζ)) = α log |1 − ζ| = α Re(log(1 − ζ)), and selecting the logarithm holomorphic in C \ (−∞, 0], for ζ = eit , |t| < π, we obtain Hu(eit ) = α · Im(log(1 − eit )) = α · arg(1 − eit ) = α · arg(eit/2 (e−it/2 − eit/2 )) ⎧ ⎪ ⎪ ⎨α(t/2 − π/2) if 0 < t < π, =⎪ ⎪ ⎩α(π/2 + t/2) if − π < t < 0. It follows that Hu∞ = |α|π/2. For the weight w(eit ) = |t|α the property results from the fact that given two weights w and W, the identity mapping f −→ f is an isomorphism L2 (wm) → L2 (Wm) if and only if the weights are equivalent, i.e. cw ≤ W ≤ Cw, where c > 0, C > 0 are constants. This implies that E = (zk )k∈Z is simultaneously a basis or not in L2 (wm) and L2 (Wm). By setting w = |1 − eit |α , W(eit ) = |t|α (|t| < π), we obtain the following conclusion. Conclusion E = (zk )k∈Z is a Schauder basis of the space L2 (wα m), as well as of L2 (|t|α m), if and only if |α| < 1.

4.7 Prediction and Hankel Operators We have already applied the techniques of Hardy spaces to stationary processes. The technique of this chapter was also in large part motivated by the needs of the theory of processes and of signal processing. Once established, it resolved a number of problems of prediction. In reality, the appropriate technique to treat the degree of mixing and ergodic properties of a process, such as regularity and singularity, is the use of the Hankel operators: see § 4.7.2 below, or sources such as Nikolski (1986).

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

126

An Introduction to Weighted Fourier Analysis

4.7.1 Strongly Regular Processes Let (xn )n∈Z be a stationary sequence in a Hilbert space H. Its past at time n is Hn− = spanH (xk : k < n) (see Definition 1.6.1), with its “past” being H− = H0− = spanH (xn : n < 0). Similarly, Hn is its future at time n and and H+ the “future” of the process, where Hn = spanH (xk : k ≥ n),

H+ = H0 .

We present below a number of properties that distinguish the set of regular processes (see Definition 1.6.1 for the definition and Corollary 2.7.2 for a spectral description) using the techniques developed in this chapter. Recall that given a subspace E ⊂ H, PE denotes the orthogonal projection onto E. (1) Let x ∈ H+ (a state in the future of the process). Then the optimal prediction with respect to the past H− is the vector PH− x satisfying the following property (and is well-defined by it) PH− x ∈ H− ,

distH (x, H− ) = x − PH− x.

Indeed, this is practically the definition of the orthogonal projection. (2) Definition. A stationary process (xn )n∈Z is said to be strongly regular if the optimal predictions are comparable with the state vectors themselves: x − PH− x ≥ cx

(∀x ∈ H+ ),

where c > 0 is a constant. The following theorem shows, among other things, that a strongly regular process is regular, and provides a criterion of strong regularity as a function of the spectral measure of the process. (3) Theorem (Helson and Szeg˝o, 1960) Let (xn )n∈Z be a stationary process, and μ its spectral measure. The following assertions are equivalent. (i) (ii) (iii) (iv)

(xn )n∈Z is strongly regular. PH− PH+  < 1. AH (H− , H+ ) > 0. μ = wm (μ s = 0) and w ∈ (HS ) (see Definition 4.6.2).

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.7 Prediction and Hankel Operators

127

Proof We have x2 = PH− x2 +x−PH− x2 (∀x ∈ H+ ) and hence x−PH− x ≥ cx if and only if (1 − c2 )x2 ≥ PH− x2 (∀x ∈ H+ ), which is equivalent to 1 − c2 ≥ PH− PH+ 2 . Thus, (i) ⇔ (ii). The equivalences of (ii), (iii), and (iv) were established in Theorem 4.6.1. 

4.7.2 Angular Operators and Hankel Operators As is clear from the proofs of Theorems § 4.7.1(3) and 4.6.1, everything boils down to the norm of the operator H x = PH− x,

x ∈ H+ ,

H : H+ → H− ,

for which we have cos(AH (H+ , H− )) = H. The operator H is called the angular operator between the future and the past. The majority of the ergodic properties of a stationary process (mixing properties, as well as the different forms of regularity) can be expressed as a function of the properties of H (more sophisticated than the norm, such as the singular numbers, etc.), and then translated into the language of spectral measures. Some references are given in § 4.9. In principle, the angular operator H can clearly be defined for an arbitrary pair of subspaces H− , H+ , but there is a price to pay: it will no longer be a Hankel operator, as is the case for the past and future of a process. The following theorem (4.7.1) is the “easy part” of a theorem of Nehari (1957), who established rich and profound links between the Hardy spaces and a vast theory of integral operators (such as those of Hankel and of Toeplitz), which is the subject of numerous monographs; see for example Nikolski (2002) and Peller (2003). See also the comments in § 4.9. In fact, we will only introduce here a restricted notion of a Hankel operator, by pre-supposing the existence of an operator symbol. The theorem itself will simply be a reinterpretation of a portion of Theorem 4.6.1. For our limited goal, we use the following definition of a Hankel operator. Let ϕ ∈ L∞ (T) and Hϕ f = P− (ϕ f ),

f ∈ H2,

where P− is the orthogonal projection in L2 (T) onto the subspace H−2 = L2 (T) H 2 . Hϕ is called the Hankel operator of symbol ϕ. It is easy to see that a symbol is not unique, and moreover, the following theorem holds.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

128

An Introduction to Weighted Fourier Analysis

Hermann Hankel (1839– 1873) was a German mathematician. He studied under M¨obius, and then Riemann, Weierstrass, and Kronecker. He defended his thesis at Leipzig in 1861, received his habilitation in 1862, and was named professor at the University of T¨ubingen in 1869. He worked on Grassmann’s linear algebra, on integration theory (preparing the road for measure theory) and on the transformations of functions of complex variables. His name is associated with the Hankel transformation (on the functions in a Euclidean space depending only on x), the Hankel function (i.e. Bessel functions of the third kind), and – especially – the Hankel matrices (and operators). The latter are defined as being the matrices A = {a(i, j) : 1 ≤ i, j ≤ n} whose elements are constant on the diagonals perpendicular to the principal diagonal, i.e. a(i, j) = ϕ(i + j) where ϕ is a function of a single variable. Today, these matrices and their continuous analogs are ubiquitous in several analytic disciplines and in their applications (harmonic analysis, the theory of holomorphic interpolation, optimal control, random processes, signal processing, etc.). Stricken with meningitis, Hankel died at 34 of a cerebral hemorrhage while traveling with his wife.

Theorem 4.7.1 (Nehari, 1957: the norm of a Hankel operator) Let ϕ ∈ L∞ (T) and let Hϕ be a Hankel operator with symbol ϕ. Then, Hϕ  = inf{ψ∞ : Hψ = Hϕ } = distL∞ (T) (ϕ, H ∞ ). Proof The equality with the “dist” follows from the simple observation that, for every f ∈ H 2 , g ∈ H−2 , we have  (Hϕ f, g)L2 (T) = (P− (ϕ f ), g) = (ϕ f, g) = ϕ f g dm, T

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.8 b(X) Versus ub(X)

129

and hence Hϕ  = sup{|(Hϕ f, g)| : f ∈ H 2 ,  f 2 ≤ 1; g ∈ H−2 , g2 ≤ 1} 

  = sup  ϕ f g dm : f ∈ H 2 ,  f 2 ≤ 1; g ∈ H−2 , g2 ≤ 1 . T

The rest of the reasoning follows exactly the proof of the implication (3) ⇒ (4) in Theorem 4.6.1. 

4.8 b(X) Versus ub(X) Here, we require a little more Banach-related terminology, i.e. relative to the more general framework of Banach spaces. Let X = (xα ) (α ∈ A) be a family of elements of a Banach space X indexed by a set of indices A (formally arbitrary, but in the examples we usually have A = Z, Z+ , Zn , etc.). Unless otherwise stated, X is assumed to be separable and A at most countable. A family X = (xα ) is called an unconditional basis of X if for any x ∈ X there exists a unique complex family (cα ) such that

cα xα , x = lim σ

α∈σ

where σ runs over the set of finite subsets of A ordered by inclusion. More precisely, this disordered convergence means that ∀ > 0, ∃σ ⊂ A (finite) such that ∀σ (finite), σ ⊃ σ



we have x − cα xα < .

(4.1)

α∈σ

Clearly an unconditional basis is a Schauder basis with respect to any numbering of A = (α j ) j≥1 (and the result of the summation does not depend  on the choice of the numbering, x = limn→∞ nj=1 cα j xα j ); in particular, an unconditional basis X = (xα ) is minimal. If X = (xα ) is a dual family of functionals, xα , xβ  = δαβ , then, by Banach’s theorem § 4.1.2(b), the projections Pσ of the partial sums,

x, xα xα , x ∈ X, Pσ x = α∈σ

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

130

An Introduction to Weighted Fourier Analysis

are continuous, and ub(X) := sup Pσ  < ∞. σ

The number ub(X) is called the unconditional basis constant. When choosing a particular numbering A = (α j ) j≥1 , we have also a corresponding basis constant b(X) (see § 4.1.1(e) for the definition). Clearly, b(X) ≤ ub(X). Also, an analog of the lemma in § 4.1.2(c) holds, and thus conversely, a minimal family X with ub(X) < ∞ is an unconditional basis. In this section we consider finite bases Xn = (x j )nj=1 (hence, simply, the free finite sequences) for large n and we examine the question of the possible upper bounding of ub(X) in terms of b(X). By a change of notation if necessary, we can suppose that X = span(Xn ). The question is important for the application of bases to approximation theory, and it is not banal even in Hilbert spaces. Note, in passing, that the orthogonal bases are characterized by b(X) = 1 and/or ub(X) = 1. Moreover, the links between b(X) and ub(X) in an arbitrary Banach space, without detailing its geometric nature, can be quite simple. Theorem 4.8.1 (1) Let Xn = (x j )nj=1 be a basis in a Banach space X. Then ub(Xn ) ≤ n m(Xn ) ≤ n b(Xn ). (2) The upper bound of (1) is sharp up to a numerical constant: there exists a free sequence Xn = (x j )nj=1 in a Banach space such that ub(Xn ) ≥ b(Xn )n/2. Proof (1) is immediate since by definition b(Xn ) ≥ m(Xn ) = max Pk,k , where (recall) Pk,l x =

l

x, xj x j ,

x ∈ Xn := span(x j : 1 ≤ j ≤ n),

j=k

and b(Xn ) = max1≤k≤l≤n Pk,l  (the norm of Pk,l and Pσ as always represent the norm of the restrictions Pk,l | Xn and Pσ | Xn ). It follows that for any  σ ⊂ {1, 2, . . . , n}, Pσ  =  j∈σ Pk,k  ≤ m(Xn ) card(σ) ≤ n m(Xn ).

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.8 b(X) Versus ub(X)

131

(2) Let X = Cn be equipped with the norm y =

n−1

|y j − y j+1 |,

j=0

where y = (y j )nj=1 ∈ Cn and y0 = 0. Let Xn = (x j )nj=1 be the natural 0 − 1 basis of Cn , x j = (δi j )ni=1 . For every k (1 ≤ k ≤ n) we have yk = k j=1 (y j − y j−1 ) and hence Pk,l y = (0, . . . , yk , . . . , yl , 0, . . . , 0) = |yk | +

l−1

|y j − y j+1 | + |yl |

j=k

≤

k

j=1

|y j − y j−1 | +

l−1

j=k

|y j − y j+1 | +

l

|y j − y j−1 | ≤ 2y,

j=1

so that Pk,l  ≤ 2 and b(Xn ) ≤ 2; moreover, if 1 = (1, 1, . . . , 1) then 1 = 1, but Podd 1 = n, hence ub(Xn ) ≥ n.  For a finite-dimensional Hilbert space the situation is much more interesting: in an unexpected manner the upper bound of Theorem 4.8.1(1) can be considerably improved; indeed ub(Xn ) always increases with sublinear growth as n → ∞, as is shown by the following theorem. Below, in Lemma 4.8.5, we shall see that this new upper bound is sharp in a sense. Theorem 4.8.2 (McCarthy and Schwartz, 1965) Let Xn = (x j )nj=1 be a basis in a Hilbert space H. Then 2

ub(Xn ) ≤ 2m(Xn )n1−(0,32)/(b(Xn ) ) .

Jacob Schwartz (1930–2009) was an American mathematician and computer scientist. He conducted research in operator theory, von Neumann algebras, parallel computing (where he was one of the pioneers), the creation of programming languages, the theory of programming, and other domains of pure and applied mathematics. In pure mathematics, he is known as the co-author (with Nelson Dunford) of a fundamental text, Linear Operators I–III (1958–1971), of roughly 2 500 pages – a turning point in the development of twentieth-century functional analysis in terms of its rigor, its breadth, and its universality. Among his other important achievements, we find the above-stated classical result on bases

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

132

An Introduction to Weighted Fourier Analysis

in Hilbert spaces (see Theorem 4.8.2). In computer science, Schwartz was the creator of the programming language SETL and of the NYU (New York University) Ultracomputer, and he was the founder (and director for 15 years) of NYU’s Department of Computer Science. After submitting his thesis at Yale University (1952, supervised by Dunford), Schwartz obtained a position as professor at the Courant Institute of Mathematical Sciences, NYU (1958–2000), and was elected to the US National Academy of Sciences (1976) and National Academy of Engineering (2000). He published 18 monographs on a variety of subjects, in pure mathematics (such as von Neumann algebras), computer science, engineering, mathematical biology, etc., not to mention around 100 research articles of the highest quality. His passing did not go unnoticed by the public at large: he merited an obituary in the New York Times (published March 3, 2009). Soon after, at a celebration of his career, his sister Judith (Dunford) said: “The intelligence [of J.S. was] so great that it seemed to weigh his head down, his omnivorous interests, his kindness and generosity – a stunning generosity that not only did not look for thanks but was puzzled and even annoyed when they came.” We begin with a lemma. To simplify the notation, we write Pk = Pk,k = (·, xk )xk . Lemma 4.8.3 Let Xn = (x j )nj=1 be a basis in a Hilbert space H, with dim H = n. Then: (1) P∗k = (·, xk )xk , P∗k Pk = (·, xk )xk 2 xk , Pk P∗k = (·, xk )xk 2 xk . (2) The operator n

P∗k Pk B= k=1

is positive Hermitian ((Bx, x) > 0, ∀x  0), hence invertible, n

Pk P∗k B−1 = , Pk 2 k=1 and for every k, BPk = P∗k B. (3) There exists a positive square root of B, 

1/2 n P∗k Pk , A = B1/2 = k=1

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.8 b(X) Versus ub(X)

133

∗

and for every k, j: Pk P j = δk j Pk , Pk = Pk where Pk = APk A−1 (and hence (Axk ) is an orthogonal basis of H). (4) ub(Xn ) ≤ A−1  · A. Proof (1) Evident computation.  (2) For any x ∈ H, we have (Bx, x) = nk=1 Pk x2 ≥ 0, and if (Bx, x) = 0, then Pk x = 0 for every k, hence x = 0 (Xn is a basis). Using (1) gives  

n n

Pk P∗k Pk −2 (Bx, xk )xk 2 xk Bx = 2 P  k k=1 k=1 =

n

Pk −2 (x, xj )x j 2 (xj , xk )xk 2 xk

j,k=1

=

n

Pk −2 (x, xk )xk 2 xk 2 xk =

k=1

n

(x, xk )xk = x,

k=1

and the result follows (B is also left-invertible, since it is self-adjoint). For the last identity, for every j, we have BP j =

n

P∗k Pk P j

k=1

=

P∗j P j

=

n

P∗j P∗k Pk = P∗j B.

k=1

(3) As B is positive definite, it admits an orthogonal basis of eigenvectors: √ Bek = λk ek , λk > 0. Clearly Aek = λk ek defines a square root of B, and A∗ = A. Thus ∗

∗

A−1 Pk A = A−2 P∗k A2 = B−1 P∗k B = B−1 BPk = Pk , hence Pk = Pk ; Pk P j = APk A−1 AP j A−1 = δk j APk A−1 = δk j Pk . (4) By (3), for any complex numbers ck , we have



2 2 2 n n n

−1 −1 2 ck Pk x = A ck Pk Ax ≤ A  ck Pk Ax k=1

k=1

= A−1 2

k=1

n

|ck |2 Pk Ax2 ≤ A−1 2 (max |ck |2 )

k=1

k

n

Pk Ax2

k=1

= A−1 2 (max |ck |2 )Ax2 ≤ A−1 2 (max |ck |2 )A2 x2 . k

k

By choosing ck = 1 if k ∈ σ and ck = 0 otherwise, we obtain Pσ  ≤ A−1 ·A for every σ ⊂ {1, 2, . . . , n}, and the result follows.  The following fact is a crucial observation because it allows control of the upper bound of B = B(Xn ) as a function of the lower bound of B = B(Xn ).

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

134

An Introduction to Weighted Fourier Analysis

 Corollary 4.8.4 Let Xn = (x j )nj=1 be a basis such that nk=1 Pk P∗k ≥ αI, where α > 0. Then n

1 m(Xn )2 I, P∗k Pk ≤ B= I= 2 α δ(Xn ) α k=1 where δ(Xn ) is the uniform minimality constant (for the equality δ(Xn )−1 = maxk Pk , see § 4.1.1(d)). Indeed, δ = δ(Xn ) = δ(Xn ) and by Lemma 4.8.3(2) αδ(Xn )2 I ≤

n

Pk P∗k = B−1 . 2 P  k k=1



Lemma 4.8.5 Let Xn = (x j )nj=1 be a basis and n = 2m . Then B=

n

P∗k Pk ≥

k=1 −1

and hence B

 1 1 + (2b(Xn ))−2 m I := αI, n

≤ (1/α)I.

Proof First note that if P is a bounded projection on a Hilbert space H and P = I − P, then P = P  (Exercise 4.9.3(a)) and (P − P )−1 = P − P , hence (P − P )−1  ≤ 2P and, for any x ∈ H, (P − P )x ≥ (2P)−1 x. The parallelogram identity 2(Px2 + P x2 ) = (P + P )x2 + (P − P )x2 = x2 + (P − P )x2 implies that for any x ∈ H we have 1 (1 + (2P)−2 )x2 . 2 We apply this lower estimate successively for P = P1,n/2 , P = P1,n/4 , P = Pn/2+1,3n/4 , etc. and use Pk,l  ≤ b(Xn ) to obtain Px2 + P x2 ≥

m−1 2

2

2 2m ≥ 1 (1 + (2b(X ))−2 )x2 , P x + P x k k n 2 m−1 1

2

and then m−2 m−1 2

2 2

2 2m−1

2 +2m−2 + P x + P x + P x k k k

k=1

k=2m−2 +1

k=2m−1 +1

≥

2m

k=2m−1 +2m−2 +1

2 Pk x

1 (1 + (2b(Xn ))−2 )2 x2 , 22

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.8 b(X) Versus ub(X)

135

etc., by repeating m times, P1 x2 + P2 x2 + · · · + P2m x2 ≥ or

1 (1 + (2b(Xn ))−2 )m x2 , 2m

 1 P∗k Pk x, x ≥ m (1 + (2b(Xn ))−2 )m x2 . 2 k=1



n



4.8.1 Proof of Theorem 4.8.2 Let l be such that n ≤ 2l < 2n and let X2l be a family X2l = Xn ∪ X obtained by adding to Xn an orthonormal basis X in a space orthogonal to H. Then, clearly m(X2l ) = m(Xn ), b(X2l ) = b(Xn ), ub(X2l ) = ub(Xn ). By Lemma 4.8.5, B(X2l ) ≥ 21l (1 + (2b(X2l ))−2 )l I := αI and B(X2l )−1 ≤ (1/α)I. Since b(X2l ) = b(X2l ) and m(X2l ) = m(X2l ), we also have B(X2l ) ≥ αI and hence, by Corollary 4.8.4: B(X2l ) ≤ (m(X2l )2 /α)I. Now, by Lemma 4.8.3(4), we have ub(X2l ) ≤ B(X2l )1/2 B(X2l )−1 1/2 ≤ m(X2l )/α = m(X2l ) 2l (1 + (2b(X2l ))−2 )−l , and then ub(Xn ) ≤ m(Xn ) 2l (1 + (2b(Xn ))−2 )l ≤ 2m(Xn )n(1 + (2b(Xn ))−2 )− ln(n)/ ln 2 2

≤ 2m(Xn ) n1−(0,32)/((b(Xn )) ) , since b(Xn ) ≥ 1 and for 0 < t ≤ 1/4, ln(1 + t) ≥ 4t · ln(5/4) (the function  t → t−1 ln(1 + t) is decreasing), and finally ln(5/4)/ ln 2 > 0, 32.

4.8.2 Gram Matrices When studying the approximation properties of sequences Xn = (x j )nj=1 in a Hilbert space H, we often use the corresponding Gram matrix   G = (x j , xk )H 1≤ j,k≤n .

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

136

An Introduction to Weighted Fourier Analysis

Jørgen Gram (1850–1916), a Danish mathematician, is known for his research in probability and statistics, numerical analysis, and number theory (Euler zeta function). He obtained his Master’s degree in 1873 (University of Copenhagen) and submitted his doctoral thesis in 1879. Research in mathematics never became his profession; however, as an amateur mathematician he became a member of the Danish Academy of Sciences (1888), received the Gold Medal of the Academy in 1885 (for his results on the zeta function), and was an editor of Tidsskrift for Mathematik (1883–1889). Gram’s professional career was spent in insurance companies: he started as an assistant and climbed the ladder to the post of company president; in 1884 he founded his own company, Skjold, and in 1910 became President of the Danish Insurance Council. We know relatively little about him other than his mathematical and professional activities, except that Gram was very active in research in forestry, long before his German colleagues developed mathematical models for the exploitation of forests. Gram is principally known for the Gram–Schmidt orthogonalization procedure (Gram (1883) and Schmidt (1907), who refers back to Gram), the Gram determinants, and the Gram matrix (see § 4.8.2). (Later, it was found that the orthogonalization procedure was already known by Laplace in 1816 and by Cauchy in 1836.) He was killed in 1916, run over by a cyclist as he was walking to a meeting at the Danish Academy of Sciences. The result of the Gram–Schmidt orthogonalization procedure can be expressed using the Gram determinants D j (D0 = 1):   v1 , v1   v , v  1 2 1  .. uj =  . D j−1  v1 , v j−1   v1  v1 , v1  v , v  1 2 D j =  .  ..  v1 , v j 

v2 , v1  v2 , v2  .. .

... ... .. .

v2 , v j−1  . . . v2 ...

v2 , v1  v2 , v2  .. .

... ... .. .

v2 , v j 

...

 v j , v1   v j , v2  ..  . .   v j , v j 

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

     ,  v j , v j−1   vj  v j , v1  v j , v2  .. .

4.8 b(X) Versus ub(X)

137

Here are some properties of the Gram matrices and their link with the techniques used in the proof of Theorem 4.8.2 (operators B and A). Up to a change of notation if necessary, we can always suppose that H = span(Xn ).    (1) For a, b ∈ Cn , we have (Ga, b)Cn = j a j x j , k bk xk H , hence G is positive, and Xn is a basis (free in H) if and only if G is positive definite. Conversely, it is easy to see that any positive definite matrix is a Gram matrix of a free sequence (hence of a basis). It suffices to take x j = G1/2 e j where (e j ) is the canonical basis of Cn . (2) Given two sequences Xn = (x j )n1 , Yn = (y j )nj , a linear operator U : H → H defined by U x j = y j (∀ j) is unitary (i.e. (U x, Uy) = (x, y), ∀x, y ∈ H) if and only if G(Xn ) = G(Yn ). (3) By defining a linear mapping T : Cn → H by T e j = x j,

1 ≤ j ≤ n,

where (e j ) is the canonical basis of Cn , we obtain (T a, T b)H = (Ga, b)Cn , hence T ∗ T = G. The operator T −1 is said to be the “orthogonalizer” (for obvious reasons). The equations δ jk = (x j , xk ) = (T e j , xk ) = (e j , T ∗ xk ) show that, for every k, T ∗ xk = ek , hence T ∗ is an orthogonalizer of the dual sequence Xn . Moreover, for every k, we have T T ∗ xk = T ek = xk , then (T T ∗ xk , xj ) = (T ∗ xk , T ∗ xj ) = (ek , e j ) = δ jk , thus the sequence (T T ∗ )1/2 Xn is an orthonormal basis of H. Conclusion (T T ∗ )1/2 is also an orthogonalizer of Xn . (4) By supposing that Xn is a normalized sequence, x j  = 1 (∀ j) (which does not change b(Xn ), or ub(Xn ), or the other geometrical properties of Xn ), by the definition in Lemma 4.8.3(1), we have Bx j = xj (∀ j), and hence B−1 = T T ∗ . Next, as the operators T T ∗ and T ∗ T are unitarily equivalent (Appendix E), so are G and B−1 , and hence ub(Xn ) ≤ G1/2 G−1 1/2 .

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

138

An Introduction to Weighted Fourier Analysis

(5) To conclude, note that the operator B1/2 is also an orthogonalizer (already mentioned in Lemma 4.8.3(3)): (B1/2 x j , B1/2 xk ) = (Bx j , xk ) = (xj , xk )x j 2 = δ jk x j 2 .

Remark 4.8.6 (sharpness of the inequality of Theorem 4.8.2) It is important to note that Theorem 4.8.2 implies a fairly unusual upper estimate: given an (infinite) basis X = (x j ) in a Hilbert space, we always have a sublinear estimation on the growth of ub(Xn ): ub(Xn ) ≤ Cn1− where > 0 depends only on the “quality of X as a Schauder basis” ( ≥ constant /b(X)2 ). Below we will see that any > 0 is realizable: for every > 0 there exists a weight w ∈ (HS ) such that for the exponential basis E = (zk )k≥0 in the space L2 (T, w) we have ub(En ) ≥ cn1−

where > 0 is of the order of const/b(E). This result of Spijker et al. (2003) will be obtained as the consequence of a series of exercises (§ 4.9.4) concerning the Helson–Szeg˝o Theorem 4.6.1.

4.9 Exercises We systematically use the notation of this chapter, in particular the Radon– Nikodym decomposition μ = μ s + wm (w ∈ L1 (T)) of a positive measure on T, E = (zk )k∈Z for the family of exponentials, etc.

4.9.1 Criterion of Linear Dependence of Exponentials Show that E = (zk )k∈Z is linearly dependent (not free) in L p (T, μ) if and only if the closed support supp(μ) is finite (⇔ μ is a finite sum of Dirac measures:  μ = Nj=1 c j δλ j , |λ j | = 1). Solution: E is not free in L p (T, μ) if and only if there exists a linear combination of   the zk , and hence a polynomial f = nk=m ak zk , an  0, such that  f  pp = T | f | p dμ = 0. The last equation is equivalent to saying that | f | p μ = 0, meaning supp(μ) ⊂ Z( f ) where Z( f ) is the set of zeros of f on T; this set is finite. 

4.9.2 Multipliers Versus Bases Let X = (xα ) (α ∈ A) be a family of elements of a Banach space X indexed by a set of indices A (formally arbitrary but in the examples we usually have

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.9 Exercises

139

A = Z, Z+ , Zn , etc.). Unless otherwise specified, X is assumed to be separable and A is at most countable. A family of complex numbers (λα ) is said to be a multiplier of X if the mapping T defined by T xα = λα xα ,

α ∈ A,

can be extended to a bounded linear operator spanX (X) → spanX (X). We denote this extension by the same letter T and also write λα = λα (T ); hence xα are the eigenvectors of T corresponding to the eigenvalues λα . We systematically identify T and (λα (T )) and assume (up to a change of notation) that spanX (X) = X. The set of the multipliers of X is denoted Mult(X): Mult(X) = {(λα ) : ∃T bounded and linear on X such that T xα = λα xα , ∀α}. Equipped with the operator norm (λα )Mult := T , Mult(X) becomes a normed space. In what follows, we use the definitions and notation found at the beginning of § 4.8: Pσ , b(X), ub(X), etc. (a) Show that Mult(X) is a Banach algebra (isometrically isomorphic to a sub-algebra of L(X), the algebra of bounded operators on X). Solution: Clear by the definitions.



(b) Show that Mult(X) ⊂ l∞ (A). Equality holds if and only if (xα ) is an unconditional basis of X. Solution: As λα (T ) is an eigenvalue of T , we have |λα (T )| ≤ T , and hence (λα ) ∈ l∞ (A). In the case of equality, Mult(X) = l∞ (A), to each λ = (λα ) ∈ l∞ (A) corresponds an operator T λ ∈ Mult(X) ⊂ L(X) and the mapping j defined by j(λ) = T λ , j : l∞ (A) → L(X) is linear and closed, since T −→ (λα (T )) is bounded. Hence by the closed graph theorem (see Appendix E) there exists C > 0 such that T λ  ≤ Cλl∞ (A) . In particular, for every finite subset σ ⊂ A, we have   T χσ = Pσ (where χσ is the characteristic function of σ), Pσ ( α∈A cα xα ) = α∈σ cα xα , and Pσ  ≤ C; this implies that X is minimal (set σ = {α}, α ∈ A), and thus, by an analog of § 4.1.2(c) already mentioned in § 4.8, X is an unconditional basis. 

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

140

An Introduction to Weighted Fourier Analysis

(c) Let μ = μ s + wm be a measure on T and E = (zk )k∈Z the exponentials in the space L2 (μ). Show that E is an unconditional basis if and only if μ s = 0 and w ≈ 1, i.e. w±1 ∈ L∞ (T). Hint Use (b). Solution: The sufficiency is clear, since if w±1 ∈ L∞ (T), then L2 (μ) = L2 (T) (with equivalence of norms). For the necessity, we use (b): every λ ∈ l∞ (Z) is a multiplier n T λ and T λ  ≤ Cλl∞ (Z) . Let ζ ∈ T and λ = (ζ )n∈Z , then the corresponding multiplier  T λ is a rotation: for any polynomial p ∈ P, p = n∈Z cn zn , we have (T λ p)(z) =  n 2 n∈Z cn (ζz) = p(ζz), and moreover T λ pL2 (μ) ≤ CpL2 (μ) . Since P is dense in L (μ), we obtain   T

| f (ζz)|2 dμ ≤ C 2

T

| f (z)|2 dμ,

∀ f ∈ L2 (μ),

which implies (with f = χE , E ⊂ T measurable) μ(ζE) ≤ Cμ(E) for all ζ ∈ T, and then (replacing E with ζE) μ(E) ≤ Cμ(ζE). By integrating over dm(ζ), we obtain C −1 μ(E) ≤ (μ ∗ m)(E) ≤ Cμ(E) for any measurable E; however, μ ∗ m = cm where c is a constant, and the result follows. 

(d) Let μ be a measure on T and let E = (zk )k∈Z be the exponentials in the space L2 (μ); use the abbreviated notation Mult(E, L2 (μ)) = Mult(μ) and T λ μ for the norm of the multiplier T λ in the space L2 (μ). Show that for two arbitrary measures μ and ν we have Mult(μ) ⊂ Mult(μ ∗ ν) and T λ μ∗ν ≤ T λ μ

(∀λ ∈ Mult(μ)).

Solution: Let f ∈ P, fζ (z) = f (ζz), T = (tk ) ∈ Mult(μ), then T fζ = (T f )ζ and hence T f 2μ∗ν    = |T f (z)|2 d(μ ∗ ν) = |T f (ζz)|2 dμ(z) dν(ζ) T T T    2 = dν(ζ) |(T f )ζ (z)| dμ(z) = dν(ζ) |(T fζ )(z)|2 dμ(z) T T T T  = T fζ 2μ dν(ζ) ≤ T 2μ  fζ 2μ dν(ζ) T T   2 2 = T μ dν(ζ) | f (ζz)| dμ(z) T T   2 = T μ | f (ζz)|2 dμ(z) dν(ζ) = T 2μ | f (z)|2 d(μ ∗ ν)(z) T

T

T

= T 2μ  f 2μ∗ν . Thus T ∈ Mult(w ∗ ν) and T λ μ∗ν ≤ T λ μ .

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press



4.9 Exercises

141

(e) Deduce that if E = (zk )k∈Z is a Schauder basis in the space L2 (μ) with a basis constant b(Eμ ) (see § 4.1.1(e) for the definition) and ν is a positive measure on T, then E = (zk )k∈Z is a basis in L2 (μ ∗ ν) with b(Eμ∗ν ) ≤ b(Eμ ). Solution: Clear by (d) and Lemma 4.4.3, because a partial sum projection Pk,l is a  multiplier, hence Pk,l μ∗ν ≤ Pk,l μ .

(f) Deduce that w ∈ (HS ) ⇒ w ∗ ν ∈ (HS ), for any positive measure ν. Solution: Clear by (e) and Theorem 4.6.1.



4.9.3 Projections on a Hilbert Space Let H be a Hilbert space and let P be a bounded projection on H (P2 = P). (a) Show that P = 1 − P. Show with an example that on a Banach space we can have P  1 − P. Solution: Denote L = PH and M = (I − P)H so that P = PLM , I − P = P ML . Then, by § 4.3.1(c), PLM −2 = sin2 (A(L, M)) = 1 − cos2 (A(L, M)). Clearly cos(A(L, M)) = cos(A(M, L)), hence PLM  = P ML . For the example in a Banach space, let X = C[0, 1] be the space of continuous functions on [0, 1] equipped with the uniform 1 norm  f ∞ = max[0,1] | f |, and P f = 0 f (x) dx. Then P2 = P, P = 1 but I − P = 1 + P = 2: choosing a function fn continuous and piecewise linear such that fn (0) = 1 and fn (x) = −1 for 1/n ≤ x ≤ 1 we obtain (I − P) fn = fn + (1 − 1/n), hence  (I − P) fn ∞ ≥ fn (0) + 1 − 1/n = 2 − 1/n and  fn ∞ = 1.

(b) Let H = PH− |(PH) , where H− = (I − P)H. Show that H2 + 1/P2 = 1. Solution: Indeed, by § 4.7.2 we have H = cos(A(PH, (I − P)H)) and by § 4.3.1(c) 1/P = sin(A(PH, (I − P)H)). 

4.9.4 The Sharpness of the McCarthy–Schwartz Inequality We propose to show, via a concrete computation linked to Theorem 4.6.1, that the upper estimate of ub(Xn ) in terms of b(Xn ) given by Theorem 4.8.2 is sharp. We have already explained the sense of this sharpness in Remark 4.8.6. (a) Let 0 < α < 1 and

  1 − t α  . wα (t) =  1 + t

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

142

An Introduction to Weighted Fourier Analysis

Show that wα = |h|2 with an outer function h, h ∈ H 2 , such that distL∞ (T) (h/h, H ∞ ) ≤ sin(απ/2), and deduce that P+  ≤

1 sin( π2 (1

− α))

.

Hint Follow the proof of (5) ⇒ (4) of Theorem 4.6.1. Solution: First observe that

  1 − t   = Hv log(wα ) = α · log  1 + t

where v = α · arg

1−t . 1+t

Since t −→ (1 − t)/(1 + t) is a mapping of the disk D to the half-plane C+ = {Re(z) > 0}, we have v∞ = απ/2 < π/2 and, following the proof of (5) ⇒ (4) of Theorem 4.6.1 (with the same notation) we obtain wα = |h|2 , h/h = e−iv+ic . Hence the domain Ω (the crucial object for our estimations) is Ω = {z ∈ C : |z| = 1, | arg(z)| ≤ v∞ }. Define the disk D(λ, R), R = (1−δ)λ, λ > 0 such that Ω ⊂ D(λ, (1−δ)λ) (as described in the proof of Theorem 4.6.1) and R/λ = 1 − δ attains its minimum: by the proof, this implies that distL∞ (T) (h/h, H ∞ ) ≤ 1 − δ. By Pythagoras, R2 = (λ − cos(απ/2))2 + sin2 (απ/2) and (1 − δ)2 = R2 /λ2 = (1 − λ−1 cos(απ/2))2 + λ−2 sin2 (απ/2), where λ > cos(απ/2). It is easy to see that min(R2 /λ2 ) is attained for λ = 1/ cos(απ/2) and gives the value (1 − δ)2 = sin2 (απ/2), so distL∞ (T) (h/h, H ∞ ) ≤ sin(απ/2). To bound P+  use Corollary 4.6.3,  h −1/2 P+  = 1 − dist2L∞ (T) , H ∞ ≤ (1 − sin2 (απ/2))−1/2 h 1 . = sin( π2 (1 − α)) 

(b) Let En = (zk )n−1 0 , n > 1, the basis of exponentials in the space Pα,n = span(zk : 0 ≤ k < n) of the polynomials of degree < n equipped with the L2 (T, wα m) norm. Deduce from (a) that 2 . b(En ) ≤ π sin( 2 (1 − α))

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.9 Exercises

143

Solution: By Lemma 4.4.3, b(En ) ≤ 2P+ , and the result follows.



(c) (Spijker, Tracogna, and Welfert, 2003). Re-using the notation of (b), let Rn : Pa,n → Pa,n be the rotation through the angle π, (Rn p)(z) = p(−z), z ∈ T. Show that, for every n, 1 1−

1 nα Rn  ≥ = n , where 0 < ≤ 2/b(En ) , 2 11 11 with c > 0, C > 0 constants.  Hint Select p = 0≤k 0) where x j (t) = e2πi jt/a , j ∈ Z. For non-self-adjoint operators, the eigenvectors are no longer orthogonal and the generalized Fourier series appear. We could again handle the situation with the help of orthogonal series of eigenvectors of a “neighboring” self-adjoint operator, but this becomes ever more complicated, and we end up turning to other developments and transformations of Fourier type (wavelets, etc.). In this chapter we have only developed the very beginning of the theory, but invite the interested reader to continue with more advanced texts such as those of Kenig (1994), Duoandikoetxea (2001), Stein (1993), Meyer (1992), and Kahane and Lemari´e-Rieusset (1998). For § 4.1–§ 4.3 we refer the reader to the great classics of functional analysis such as Banach (1932), Riesz and Sz.-Nagy (1955), and Lindenstrauss and Tzafriri (1977). The angle between two subspaces of a Hilbert space was

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

146

An Introduction to Weighted Fourier Analysis

defined by Friedrichs (1937), when developing a geometric approach to the problems of perturbation of operators. Section 4.4 is found, for the most part, in Kolmogorov (1941), but we have taken into account important later developments: Helson and Szeg˝o (1960), Helson and Sarason (1967), Ibragimov and Rozanov (1970), and Peller and Khruschev (1982). For comments on the Hilbert operator H (§ 4.5) see Notes and Remarks 2.9. We can also add that Hilbert himself defined the operation P+ (which is now known as the Hilbert transform, or the Riesz projection) in his famous course on integral equations (1904–1908 (Hilbert, 1912)) where he made progress on the Riemann problem dealing with the basic equation in singular integrals (Cauchy kernel). Today, the techniques linked to the Riemann– Hilbert problem are crucial tools in a dozen mathematical disciplines: from complex analysis, integral operators, and integrable models in mathematical physics through to probabilistic combinatorics, not to mention orthogonal polynomials, random matrices, and signal theory. It is not surprising that throughout the twentieth century this field attracted such enormous attention; in pure analysis, it culminated in the theory of the Calder´on–Zygmund integral operators. The treatment in § 4.6 is close to that of Nikolski (1986), but the main contents comes from Helson and Szeg˝o (1960). It is interesting to note that while Theorem 4.6.1 provides a complete characterization of whether the subspaces H 2 (w) and H−2 (w) are at a strictly positive angle to each other, the natural question “When is the projection P+ (x− + x+ ) = x+ well-defined on the vector sum H 2 (w) + H−2 (w)?”, i.e. when is H 2 (w) ∩ H−2 (w) = {0}, remains open: see Sarason (1994). We must also point out that since the appearance, thanks to Burkholder, Hunt, Muckenhoupt and Wheeden (see Hunt et al., 1973), of an approach to weighted analysis totally different from that of Helson and Szeg˝o, the two complement each other for different applications and generalizations. The Helson– Szeg˝o (HS) approach is based on the operation of harmonic conjugation which is not “local” and whose existence requires a group structure. By contrast, the Hunt–Muckenhoupt–Wheeden (HMW) approach is predominantly “local,” realizable on a more or less arbitrary metric space, and is based on the technique of scaling – an operation close to many “dyadic” techniques, e.g. wavelets. More precisely, given a weight w ≥ 0 on T, the condition (HMW), known as the Muckenhoupt condition (A2 ), which is equivalent to the condition

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.10 Notes and Remarks

147

w ∈ (HS ) of Theorem 4.6.1(5), consists of   1  1  1  dm < ∞, w dm A2 (w) := sup m(I) I m(I) I w I where I runs over the set of sub-intervals (arcs) of T. To mention only a few selected points of “competition” between these two approaches, we note the following. (i) (HS) works better at providing simple examples of weights w ∈ (HS ) = (A2 ) (beginning with the Babenko Example 4.6.6), and it is more easily adaptable to harmonic analysis in several variables (on the circle Tn , for example) and for the two-weight problem, consisting of characterizing the pairs w1 , w2 ≥ 0 such that   |H f |2 w1 dm ≤ C 2 | f |2 w2 dm, ∀ f ∈ P T

T

(see Cotlar and Sadosky, 1979). (ii) By contrast, (HMW) is indispensable in the analysis of singular integrals on manifolds without a group structure, and is better adapted for applications to multi-dimensional random processes, etc. It is quite easy to see that (HS ) ⇒ (A2 ), but there is no direct proof for the converse (A2 ) ⇒ (HS ). During the twentieth century, the applications to stationary processes (§ 4.7) and signal processing were the principal motors of development of Fourier analysis, both classical and generalized. For the period of the “youth of the theory” (which is all that is presented in this book), refer to Kolmogorov (1941) and Wiener and Masani (1957, 1958), and for subsequent developments, to Rozanov (1963) and Ibragimov and Rozanov (1970); for an innovative survey article, see Peller and Khruschev (1982). In particular, the very tight links between processes and Hankel operators were discovered in the lastcited article. The Hankel operators (see § 4.7.2 and Theorem 4.7.1, but we have not given the general definition here), thanks to the efforts of Nehari, Krein, Adamyan, Arov, Peller, and others, have been transformed into a powerful and indispensable tool for the study of a certain number of analytic phenomena such as random processes, signal processing and H ∞ optimal control, interpolation theory, theory of best approximations, etc. The references for § 4.7.2 are Nehari (1957), Peller (2003), Power (1982), and Nikolski (2002). It is interesting to note that several properties of the operators Hϕ = P− ϕ | H 2 (in principle, quite special objects closely linked to harmonic analysis on T) are shared with abstract expressions of the type H = PE ⊥ T |E where E is

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

148

An Introduction to Weighted Fourier Analysis

a subspace of a Hilbert space on which operates a bounded operator T (for Hϕ this is the operator of multiplication by ϕ); see Devinatz and Shinbrot (1969) or the monographs mentioned above. The question of the “quantitative qualities” of the bases in a vector space (and in particular, in a Hilbert space), tackled in § 4.8, is extremely important for any application to numerical analysis and matrix analysis, but also for purely theoretical problems in high-dimensional geometry. We have limited ourselves to the beautiful result of McCarthy and Schwartz (1965) (Theorem 4.8.2), and to its converse by Spijker, Tracogna, and Welfert (2003) (§ 4.9.4(c)). It is interesting to note that this result remains somewhat mysterious (despite the extreme transparency of the original proof), because it does not provide an explanation of the origin of this improvement in the order of ub(Xn )/b(Xn ) in a Hilbert space compared to a general Banach space. It is also curious to know that the authors of this analytic gem from 1965 were convinced that their result was quite approximate and that the true rate of growth of ub(Xn ) would be logarithmic; they put forward the conjecture ub(Xn ) = O((log n)c(b(Xn )) b(Xn )) as n → ∞. It is also very instructive that the result of § 4.9.4(c) (the converse of Theorem 4.8.2) was found 35 years later by specialists in applied numerical analysis (Spijker, Tracogna, and Welfert, 2003). (The calculations of Spijker et al., 2003 are different from those presented in § 4.9.4(c).) The result of § 4.9.4(c) is based on the inequality § 4.9.4(a) which is, in fact, the equality P+  = 1/ sin( π2 (1 − α)) (Hollenbeck and Verbitsky, 2000); these authors found exact values for norms of many other operators of harmonic analysis ). Another aspect of the question of the “quality of a basis” is to regard it in the form of a summation basis. More precisely, let X = (xk ), X = (xk ) be a complete and total biorthogonal pair, and let V = (vαk ) be a “matrix” of complex numbers such that

|vαk | · xk  · xk  < ∞ (∀α) and lim vαk = 1 (∀k). α

k

Then, V is said to define a summation method and the V-sum of the series   k x, xk xk is taken as the limit (if it exists)



(V) x, xk xk = lim vαk x, xk xk . k

α

k

We only consider the “hereditary” methods, i.e. where the convergence of a    series k x, xk xk implies (V) k x, xk xk = k x, xk xk . For example, the method of arithmetic means (Ces`aro and Fej´er) corresponds to vnk = Ank where Ank = max(0, 1 − |k|/(n + 1)), and that of Abel and Poisson to vrk = Prk ,

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

4.10 Notes and Remarks

149

 Prk = r|k| , 0 < r < 1 (so that, for example, (P) k≥0 (−1)k (k + 1) = 1/4, but  (A) k≥0 (−1)k (k + 1) does not exist, whereas if the (A)-sum exists then the (P)sum also, and they are equal). Rosenblum (1962) proved that E = (zk )k∈Z in L2 (wm) is a basis for the Abel–Poisson method (P) if and only if w ∈ (HS ) (hence it is already a Schauder basis). It is not known how to characterize the weights w on T for which there exists a summation method for the Fourier series in L2 (wm) (we suppose, of course, that w±1 ∈ L1 (T)). The same is true for the following property of spectral synthesis (the weakest property guaranteeing a “reconstruction” of every function f ∈ L2 (wm) from its Fourier  f (k)zk ): series k∈Z ! f (k)zk : k ∈ Z) f ∈ spanL2 (wm) ( !

(∀ f ∈ L2 (wm)).

In the literature, this property is also known as hereditary completeness, or strong M-basis. For recent news on the hereditary (non)completeness of nonharmonic exponentials and systems of reproducing kernels see Baranov, Belov, Borichev (2013) and Baranov, Yakubovich (2016) The Gram matrix techniques of § 4.8.2 are classical in this group of ideas and ubiquitous in all applied matrix analysis, as well as in approximation theory. Indeed, it is easy to see that any positive matrix ((Ax, x) ≥ 0, ∀x) is a Gram matrix of a sequence of vectors. For Gram matrices, see Golub and Van Loan (1996), Akhiezer (1965), and Gantmacher (1966). The multipliers, and especially the Fourier multipliers (i.e. for the family of exponentials E) count among the indispensable subjects of harmonic analysis, as they can be identified with the eigenvalues of the convolution operators (invariant with respect to translations). We refer to Zygmund (1959) for an introductory presentation.

https://doi.org/10.1017/9781316882108.006 Published online by Cambridge University Press

5 Harmonic Analysis and Stationary Filtering

Topics. Filters (finite-power, stable, causal), harmonic signals, time and frequency domains, transfer function, Wiener’s fundamental theorems, synthesis of filters, band-pass filters, Rudin–Carleson theorem, Helson sets, inverse problems, five differences between C and W, a brief overview of sampling. The mathematical theory of stationary filtering was founded by Wiener in the 1930s, but also by Kolmogorov, Masani, and others, and, on the engineering side, by Kotelnikov and (independently, but 15 years later) by Shannon; see Notes and Remarks 5.7 at the end of this chapter (including the biographies). Throughout the twentieth century, filtering theory amply nourished all of harmonic analysis by proposing fundamental problems to be solved. In particular, this is the case for the theory of Hardy spaces, so that separating the theoretical applications of filtering from the theory of Hardy spaces itself has today become delicate. This is why the contents of this chapter can be considered more a “filtering interpretation” of the theory already developed, rather than a new subject. More precisely, in this chapter we only consider discrete-time signals. For continuous-time signals, there are a few references given in § 5.7. Finally, we warn the reader that in the different mathematical presentations of signal processing, the terminology linked with the physical nature of the signals (such as the energy, power, etc.) can vary. Here, we follow the language of the founders of the theory (see the references in § 5.7).

5.1 The Language of Linear Filters By definition, a signal x is a complex-valued function of a variable called the time. We will study signals of discrete time Z, i.e. of complex sequences 150 https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.1 The Language of Linear Filters

151

n −→ xn , n ∈ Z, hence x = (xn )n∈Z . The principal operation on signals is filtering – which consists of passing a signal x through an apparatus Φ (a “box”) which transforms it into another signal y: Φ : x −→ Φx = y.  The energy of a signal is defined as n∈Z |xn |2 . We will only consider filters Φ that are stationary and of finite power; these are described by the following axioms. Definition 5.1.1 A finite-power stationary filter Φ is a mapping of numerical sequences (xk )k∈Z satisfying the axioms (A1 )–(A4 ): (A1 ) Φ is of finite power, i.e. it transforms a signal with finite energy into another signal with finite energy, hence Φ is a mapping of the space l2 (Z) into itself, (A2 ) Φ is linear, (A3 ) Φ is stationary (invariant by translation on Z), i.e. for every n ∈ Z Φτn = τn Φ, where τn ((xk )k∈Z ) = (xk−n )k∈Z is a translation of step n on Z, (A4 ) Φ is correctly observable, i.e. the observation of a coordinate x −→ (Φx)0 is continuous on l2 (Z). Another important class of filters is that of “stable” filters. A filter Φ is said to be stable stationary if satisfies (A2 )–(A4 ) and the following axiom (A1 ) in place of (A1 ). In (A1 ) we replace l2 (Z) by l∞ (Z): (A1 ) Φ is well-defined on the bounded signals x = (xk )k∈Z , x ∈ l∞ (Z), and transforms them into bounded signals, hence Φ is a mapping of the space l∞ (Z) into itself. Remark 5.1.2 (diagonalization) The eigenvectors of the group of translations (τn )n∈Z , i.e. the sequences x = (xk )k∈Z satisfying τn x = λn x,

n ∈ Z,

where λn ∈ C, play a particularly important role in the analysis of filters. It is easy to find them: τ1 x = λx ⇔ x = a · xλ , where xλ := (λ−k )k∈Z (here a ∈ C, λ ∈ C∗ = C \ {0}).

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

152

Harmonic Analysis and Stationary Filtering

The signal xλ := (λ−k )k∈Z is called the input harmonic signal of frequency λ (or of frequency arg(λ)). Clearly the harmonic signals (or, more briefly, the harmonics) in a way diagonalize a stationary filter: λ(Φxλ ) = Φτ1 xλ = τ1 (Φxλ ) and hence Φxλ = a(λ)xλ , for any xλ in the domain of definition of Φ; the numerical function λ −→ a(λ) is an important characteristic of the filter (see below). This is a key idea for the development of a theory of stationary filtering, but in this form, it remains somewhat heuristic. A weak point in the above reasoning is that we do not know the xλ for which Φ is well-defined: it could be that this set is simply empty (in particular, there is no xλ with finite energy, but there are cases where it is possible to directly follow the above reasoning – for example, for the stable filters). To work around this difficulty of “infinite energies,” we use the discrete Fourier transform.

5.1.1 The Fourier Transform and the Frequency Domain By the axioms (A1 )–(A2 ), a stationary filter Φ is a linear mapping Φ : l2 (Z) → l2 (Z) satisfying Φτn = τn Φ for every n ∈ Z. The Fourier transform F is defined in the space L1 (T) by F f = (! f (n))n∈Z , and the inverse Fourier transform by F −1 ((xn )n∈Z ) =



xn ζ n ,

ζ ∈ T.

n∈Z

The series is formal, however – in more precise contexts – we can find a manner in which it converges. For example, we know that F is a unitary mapping between the spaces L2 : F : L2 (T) → l2 (Z), F L2 (T) = l2 (Z), F f l2 (Z) =  f L2 (T) (∀ f ∈ L2 (T)), (theorems of Fourier–Plancherel, or of Riesz–Fischer: see Appendix A). In signal processing, the space l2 (Z) is known as the time domain, and the space L2 (T) as the frequency domain (or spectral domain). Signals with finite

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.2 Characterization of Stationary Filters

153

energy x ∈ l2 (Z) are also called pulse signals. The following diagram defines a representation of a filter in the frequency domain: Φ : l2 (Z) −−−−−−→ l2 (Z) ⏐ / ⏐ ⏐ ⏐ ⏐ ⏐ ⏐ ⏐ F ⏐ 1F −1 ⏐ ! : L2 (T) −−−−−−→ L2 (T) Φ i.e. ! = F −1 ΦF : L2 (T) → L2 (T). Φ The stationarity, i.e. the fact that Φ commutes with τn , is transformed into ! ! zn = Mzn Φ, ΦM since, for every n ∈ Z, F −1 τn F = Mzn , where Mz is the operator of multiplication by z. Now, we are able to find the temporal and frequency characterizations of the stationary filters.

5.2 Characterization of Stationary Filters The following characterizations reduce the analysis of filtering to the techniques of convolution operations. Theorem 5.2.1 (Wiener, 1933) (1) A finite-power stationary filter Φ is a bounded linear operator on l2 (Z), ! is bounded on L2 (T). and its frequency representation Φ (2) Let Φ be a finite-power stationary filter and ϕ = F −1 S , where S = Φe0 , e0 = (δ0 j ) j∈Z . Then ϕ ∈ L∞ (T), Φ = ϕ∞ and  

! f = ϕ f (∀ f ∈ L2 (T)). xk S n−k (∀x ∈ l2 (Z)), Φ Φx = x ∗ S := n∈Z

k∈Z

(3) Conversely, for any ϕ ∈ L∞ (T), the mapping Φx = x ∗ S , where S = F ϕ, is a finite-power stationary filter. The correspondence Φ ↔ ϕ is bijective. Proof (1) By the closed graph theorem (see Appendix E), it suffices to show that the mapping Φ : l2 (Z) → l2 (Z) is closed, i.e. (uk ∈ l2 (Z),

lim uk 2 = 0, k

lim Φuk = v) k

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

⇒

v = 0.

154

Harmonic Analysis and Stationary Filtering However, the functional x −→ (Φx)0 is continuous, hence so is x −→ (Φ(τn x))0 = (τn Φx)0 = (Φx)−n

(∀n ∈ Z),

hence 0 = limk (Φuk )n = (v)n for every n ∈ Z. Thus v = 0, and Φ is continuous. ! n f) = (2) By the stationarity commutation relation, for any f ∈ L2 (T), Φ(z ! f , hence Φ(p ! f ) = pΦ ! f for any polynomial p ∈ P. By taking f = zn Φ −1 ! ! = pϕ for every polynomial p ∈ P, 1 = F e0 , we obtain Φ(p) = pΦ1 2 ! f ) = f ϕ. ! where Φ1 = ϕ ∈ L (T). We show that, for every f ∈ L2 (T), Φ( 2 As ϕ ∈ L (T), it suffices to show that the Fourier coefficients coincide: for every n ∈ Z,  ! (Φ( f ))(n) = ( f ϕ)(n). Indeed, this is the case for the set of polynomials f ∈ P which is dense in L2 (T). Moreover, both sides are continuous in L2 (T): the left by (A4 ) and the right by the Cauchy–Schwarz inequality (since ϕ ∈ L2 (T)), hence ! f ) = f ϕ, for every f ∈ L2 (T). Hence, ϕ is a multiplier of L2 (T) and Φ( by Exercise 1.8.3(a) (or see Appendices A and D), ϕ ∈ L∞ (T) and Φ = ! = ϕ∞ . Φ  ! = Φe0 , is The formula (Φx)n = k∈Z xk S n−k , where S = F ϕ = F Φ1 immediate since it holds for x = e j = τ j e0 = (δk j )k∈Z : (Φτ j e0 )n = (τ j Φe0 )n = (τ j S )n = S n− j . ! f = f ϕ is bounded L2 (T) → L2 (T), and for every (3) Clearly, the mapping Φ ! −1 is a stationary filter. ! = ΦM ! zn , hence Φ = F ΦF  n ∈ Z, Mzn Φ There is an analog of Theorem 5.2.1 for stable filters. Theorem 5.2.2 (Wiener, 1933) (1) A stable stationary filter Φ is a bounded linear operator on l∞ (Z). (2) Let Φ be a stable stationary filter. Then S = Φe0 ∈ l1 (Z), e0 = (δ0 j ) j∈Z , Φ = S l1 (Z) and Φx = x ∗ S :=





(∀x ∈ l∞ (Z)).

xk S n−k

k∈Z

n∈Z

(3) Conversely, for every S ∈ l1 (Z), the mapping Φx = x ∗ S is a stable stationary filter. The correspondence Φ ↔ S is bijective.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.2 Characterization of Stationary Filters

155

Proof The proof is similar to that of Theorem 5.2.1, with the difference that F −1 l∞ (Z) is no longer a space of functions on T, but of distributions; we thus try to avoid it. (1) We use the same proof as in Theorem 5.2.1 (replacing l2 (Z) by l∞ (Z)). (2) By setting S = Φe0 and using (A3 ), we obtain Φen = τn S = S ∗ en , hence by linearity, Φc = S ∗ c for any signal with finite support (i.e. a linear  combination c = k ck ek , where supp(c) = {k : ck  0} is a finite set). Moreover, by (1), Φ < ∞, hence S ∗ c∞ ≤ Φ · c∞ . The matrix of the mapping 

 xk S n−k x −→ x ∗ S := n∈Z

k∈Z

is A = (ank ) where ank = S n−k (a Toeplitz matrix on Z, also called a Laurent matrix), thus by Lemma 5.2.4 below we obtain S ∈ l1 (Z) and Φ = S l1 (Z) . (3) Φ is bounded in l∞ (Z) by Lemma 5.2.4; Φ is stationary because τn Φx = en ∗ S ∗ x = S ∗ en ∗ x = Φτn x, according to well-known properties of convolutions (see Appendix A).



Remark 5.2.3 The program of diagonalization of a filter, described in § 5.1.1, can also be given for stable filters. Indeed, there is a family of eigenvectors xλ = (λ−k )k∈Z of the group of translations (τn ) that belong to the space l∞ (Z): xλ ∈ l∞ (Z) ⇔ |λ| = 1

⇔

λ ∈ T.

We start with the convolution operator S corresponding to a stable filter Φ by Theorem 5.2.1 (the impulse response of Φ, according to the terminology of § 5.3.1) and its Fourier transform

S k ζ k , ζ ∈ T, ϕ(ζ) = (F −1 S )(ζ) = k∈Z

the (transfer function of Φ, by § 5.3.1). Then, for any λ ∈ T, we have 

 Φxλ = λ−k S n−k = ϕ(λ)(λ−n )n∈Z = ϕ(λ)xλ . k∈Z

n∈Z

Hence, the value ϕ(λ) is an “amplifying factor” of the harmonic xλ = (λ−k )k∈Z = (e−ikθ )k∈Z by the filter Φ; here λ = eiθ : θ (or λ itself) is called the frequency of xλ .

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

156

Harmonic Analysis and Stationary Filtering

The family of harmonics xλ is weak-∗ complete in the space of the bounded signals l∞ (Z) (with respect to the weak-∗ topology, σ(l∞ , l1 ): see Appendix D): spanσ(l∞ ,l1 ) (xλ : λ ∈ T) = l∞ (Z).  Indeed, if a ∈ l1 (Z) and 0 = a, xλ  = k∈Z ak λ−k = (F −1 a)(λ−1 ) for every λ ∈ T, then a = 0 (since ak are the Fourier coefficients of F −1 a, an absolutely convergent Fourier series). Lemma 5.2.4 Let A = (a jk ) j,k∈J be a matrix on an at most countable set of indices J, and let c00 = c00 (J) be the vector space of finitely supported functions on J equipped with the norm x∞ = sup j |x j |. Suppose 

 Ax = a jk xk ∈ l∞ (J) j∈J

k∈J

for every x ∈ c00 (J). Then,

 (1) A : c00 → l∞ (J) = sup j∈J k∈J |a jk |.  (2) A : l∞ (J) → l∞ (J) = sup j∈J k∈J |a jk |. Proof (1) By definition, A : c00 → l∞ (J) = sup{Ax∞ : c ∈ c00 , x∞ ≤ 1} 

 

    = sup sup a jk xk  = sup sup  a jk xk  j x∞ ≤1 k∈J x∞ ≤1 j k∈J

|a jk |. = sup j∈J k∈J

(2) follows from (1) and the fact that |

 k∈J

a jk xk | ≤ x∞

 k∈J

|a jk |.



5.3 What Can Filtering Do? This section describes a number of problems that arise in the mathematical theory of filters.

5.3.1 A Bit More Terminology for Filters In addition to the preceding notation of Theorems 5.2.1–5.2.2, the following language is also used. – S = Φe0 is the impulse response of a stationary filter Φ. −1 – The function ϕ = F Φe0 is the transfer function, or frequency characteristic, or frequency response, or impedance, or voltage gain. For

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.3 What Can Filtering Do?

157

finite-power filters, we have ϕ ∈ L∞ (T), and for stable filters, ϕ ∈ F −1 l1 (Z) := W, the Wiener algebra of absolutely convergent Fourier series. iλ iλ – The function e −→ |ϕ(e )| is called the energy spectrum. – The number ϕ∞ is the amplitude distortion.

– A band-pass filter is defined by the condition |ϕ| = χσ , σ ⊂ T (σ is a band of frequencies that pass without distortion of the amplitudes). The ideal band-pass filter is ϕ = χσ (there is no distortion of either the amplitude or the phase on σ). – An all-pass filter (or dephasing filter, or phase correction filter) satisfies |ϕ| = 1. – A signal x = (xk )k∈Z is said to be causal (or positive time) if xk = 0 for k < 0. – A filter Φ is said to be causal, or physically realizable, if x causal ⇒ Φx causal. – Φ is said to be stable stationary if it satisfies (A2 )–(A4 ), is well-defined on the bounded signals x = (xk )k∈Z , x ∈ l∞ (Z), and transforms them into bounded signals. iλ – The phase lag at the frequency λ of a filter Φ is by definition arg(ϕ(e )). Similar terminology exists for the pulse signals x ∈ l2 (Z). −1 2 – The function F x ∈ L (T) is called the energy spectrum of x. −1 iλ 2 – |F x(e )| is called the energy density at the frequency λ (for it to be defined everywhere, it suffices to require x ∈ l1 (Z)). −1 iλ – arg(F x(e )) is a phase of x at the frequency λ.

5.3.2 Some Typical Problems in Filtering Here is a short list of problems in the mathematical theory of filters. (1) Direct problem. Construct a (“white box”) filter having a given frequency response on a particular band of frequencies (and as an option, physically realizable). In particular, describe a filter that detects a useful signal against background noise. (2) Inverse problem. Identify an unknown (“black box”) filter Φ : x −→ y from the harmonic analysis of an observable input/output couple x, y. In particular, study the possibility of reconstructing Φ when the spectral densities |F −1 x|2 and |F −1 y|2 are known.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

158

Harmonic Analysis and Stationary Filtering

(3) Problem of causality. Study questions (1)–(2) for finite-power and/or stable causal filters. We first examine problem (1).

5.4 Synthesis of Causal Filters The construction of a filter having certain desired characteristics is known as the “synthesis of a filter.” We begin with a frequency description of the causal filters. Lemma 5.4.1 (Wiener, 1930) A stationary filter Φ (finite-power, or stable) is causal if and only if its transfer function ϕ = F −1 Φe0 is in H ∞ , ϕ ∈ H∞ (in the case of a stable filter, ϕ ∈ H ∞ ∩ F −1 l1 (Z) = F −1 l1 (Z+ ) := Wa ). Proof The necessity is evident since e0 is a causal signal, and thus so is Φe0 , ϕ(k) = 0 for every k < 0. hence (Φe0 )k = ! For the sufficiency, let x = (xk )k∈Z be a causal signal (in l2 or l∞ ), xk = 0 for every k < 0. Then, for n < 0 and with S = Φe0 we have S n−k = 0 for every k ≥ 0, and thus



xk S n−k = xk S n−k = 0, (Φx)n = k∈Z

k≥0

and the result follows.



Corollary 5.4.2 Let w ∈ L∞ (T), w ≥ 0, w  0. The following assertions are equivalent. (1) There exists a finite-power causal filter with energy spectrum w. (2) log(w) ∈ L1 (T). This is evident by the theorems of Szeg˝o (Corollary 2.6.2) and Smirnov (§ 3.3.1(g)).  Corollary 5.4.3 Let Φ  0 be a causal filter and ϕ its transfer function. Then: (1) Φ is an ideal band-pass filter if and only if Φ = id, (2) Φ is a band-pass filter if and only if it is all-pass, and if and only if ϕ is an inner function, (3) Φ is stable and all-pass if and only if ϕ is a finite Blaschke product.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.4 Synthesis of Causal Filters

159

Indeed, (1) and (2) follow directly from the boundary uniqueness Theorem 1.4.4. For (3), if Φ is stable and ϕ is not reduced to a finite Blaschke product, then the spectrum σ(ϕin ) contains at least one point λ on the boundary of the disk D (see Definition 3.2.2). By the description of Corollary 3.2.4, limz→λ |ϕ(z)| = 0. However as it is in F −1 l1 (Z+ ) = Wa , the function ϕ is continuous in D, whence ϕ(λ) = 0 and limz→λ,z∈T ϕ(z) = 0, which is not compatible with the property “all-pass.” Hence, ϕ is indeed a finite Blaschke product.  Remark 5.4.4 (filter with minimal mean lag) Real-life applications require filters that allow the passage of a band of frequencies with a minimum of distortion (and suppress – or almost suppress – another band). According to the filter formula y = Φx,

(F −1 y)(t) = ϕ(t)(F −1 x)(t)

(for all t ∈ T),

we can distinguish an amplitude distortion |ϕ(t)| − 1 and a phase distortion arg(ϕ(t)). If we suppose that the question of amplitude is somehow resolved, so that the energy spectrum of the filter |ϕ| is chosen, it remains to minimize the phase lag θ, ϕ(t) := |ϕ(t)|eiθ(t) . Note that a constant lag θ = constant does not present a problem since it can be compensated by a multiple of the identity filter eiτ id. The question is to minimize the variation of the phase lag. There is a large variety of concrete situations where one criterion of optimization is chosen over another (for example, only a finite set of frequencies might be considered, etc.). However, here we present only one (linked to the techniques of Hardy spaces), namely, minimize the following weighted variation of the phase:   θ(u) − θ(v) dm(u) dm(v). |ϕ(u)| · |ϕ(v)| sin2 P(ϕ) = 2 T T With this formulation of the problem, it is possible to obtain a complete solution. Theorem 5.4.5 Let w be a causal energy spectrum, i.e. w ≥ 0, w ∈ L∞ (T) and log(w) ∈ L1 (T). Among all causal filters Φ with |ϕ| = w, the minimum of the weighted phase variation P(ϕ) is attained (exclusively) for outer transfer functions ϕ = c[w], c ∈ T, and 2     w dm − exp 2 log(w) dm . min P(ϕ) = T

T

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

160

Harmonic Analysis and Stationary Filtering

Proof Write 2   2    |ϕ| dm −  ϕ dm = (|ϕ(u)| · |ϕ(v)| − ϕ(u)ϕ(v)) dm(u) dm(v) T

T

T

T

(the integral with Im(ϕ(u)ϕ(v)) is canceled given the oddness of the sine function)   = |ϕ(u)| · |ϕ(v)|(1 − cos(θ(u) − θ(v))) dm(u) dm(v) = 2P(ϕ), T

T

hence the minimum of P(ϕ) among the ϕ ∈ H ∞ , |ϕ| = w, is attained if and only if 2   max  ϕ dm = max |ϕ(0)|2 T

is attained. However, by Jensen’s inequality in Lemma 2.3.1, we always have    |ϕ(0)|2 < exp 2 log(w) dm , T

with the exception of equality only in the case where ϕ = c[w] (see the criterion of Theorem 2.6.7). 

5.4.1 Filters of Optimal “Signal to Noise Ratio” This problem concerns a signal x of a known form that is polluted by a random parasite signal b, assumed to be independent white noise. More precisely, b(·) = (bk (·))k∈Z is a sequence of random independent variables on a probability space (Ω, dω), all with expectation 0 and variance 1. The problem is to construct a finite-power (or stable) stationary filter Φ, providing, at a fixed moment n ∈ Z, the best ratio of useful output signal yn = (Φx)n against the quadratic mean of the noise, i.e. giving max

|(Φx)n | |yn | = max   . B |(Φb(ω))n |2 dω 1/2 Ω

A filter Φ providing the maximum, if such exists, is said to be optimal. The following theorem provides a solution. Theorem 5.4.6 (1) Given x, b and n, an optimal finite-power (respectively, stable) filter exists if and only if F −1 x ∈ L∞ (T), hence x is a signal of bounded energy density (respectively, x ∈ l1 (Z)).

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.4 Synthesis of Causal Filters

161

(2) Suppose F −1 x ∈ L∞ (T) (respectively, x ∈ l1 (Z)). Then the only filter Φ giving max(|yn |/B) is the filter with impulse response S k = (Φe0 )k = cxn−k (k ∈ Z) where c  0 is a constant (an “adapted filter”).  Proof Given a finite-power filter Φ, we have (Φb(ω))n = k∈Z bk (ω)S n−k , where the functions ω −→ bk (ω), ω ∈ Ω, form an orthonormal sequence, and hence

= S l2 (Z) . B = bk (·)S n−k L2 (Ω) 

k∈Z

 However, yn = k∈Z xk S n−k = k∈Z xn− j S j = (S , x)l2 , where x = (xn− j ) j∈Z . Consequently, # " " # |(S , x)l2 | |yn | ∞ : Φ of finite power = sup sup : S ∈ F L (T) = xl2 (Z) , B S l2 (Z) since the image F L∞ (T) is clearly dense in l2 (Z) (along with others, it contains the sequences with finite support c00 ). It is well-known that the Cauchy–Schwarz inequality |(S , x)l2 | < S l2 (Z) xl2 (Z) is strict with only a single exception, S = cx where c  0 is a constant. Hence the sup is attained with a finite-power filter if and only if F −1 x ∈ L∞ (T), in accordance with the statement of the theorem. Clearly the version for stable filters is also resolved. 

5.4.2 Frequency Response on a Very Thin Band Let σ ⊂ T be a Borel set (a band of frequencies). Is it possible to find a finite-power or stable causal filter Φ having a transfer function ϕ equal to an arbitrary bounded (or continuous) function on σ? An evident restriction is given by the uniqueness theorem of Corollary 1.4.4: necessarily, m(σ) = 0 (otherwise, by decomposing σ = σ1 ∪ σ2 where σ j are disjoint and m(σ j ) > 0, j = 1, 2, it would not be possible to find ϕ ∈ H ∞ satisfying ϕ|σ = χσ1 ). Another fact – not really a restriction but merely an inconvenience – is that the transfer function of a generic finite-power filter is defined almost everywhere on T and not everywhere. The outcome of this somewhat ambiguous situation consists in regarding σ = σ,

m(σ) = 0

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

162

Harmonic Analysis and Stationary Filtering

and the causal filters Φ having a continuous transfer function ϕ (if Φ is stable, ϕ is automatically continuous), i.e. ϕ ∈ Ca (D) = H ∞ ∩ C(T). Other than these two conditions, there are no other constraints when constructing a finite-power filter having a predetermined frequency response on σ. It turns out that the response for stable filters is completely different: to have an arbitrary continuous frequency response on σ, it is necessary that σ be a “very thin” set with a very specific arithmetical structure. This kind of set σ, such that C(σ) = Wa | σ, is called a Helson set (Helson introduced them in 1954). Even today, there does not exist any intelligible description of what makes a set a Helson set. We give below a few examples of Helson and non-Helson sets. We begin with the following theorem, proved independently by Walter Rudin (also known for his university textbooks in mathematics) and Lennart Carleson, indicating that on the sets of measure zero, the spaces Ca (D) and C(T) are indistinguishable.

Walter Rudin (1921–2010) was an American mathematician, one of the primary experts in the harmonic and complex analysis of the years 1950– 1990. He published a series of books (university textbooks and research monographs) of unequaled mathematical and pedagogical quality, whose influence on the teaching of mathematical analysis and the formation of the new generations of mathematicians worldwide is incontestable. The most well known are Real and Complex Analysis (1966), Functional Analysis (1973), Principles of Mathematical Analysis (1953) (nicknamed “Baby Rudin”) and also Fourier Analysis on Groups (1962). Rudin was rewarded with the Steele Prize for Mathematical Exposition in 1993. He came from a well-known European Jewish family that had lived in Vienna for centuries. His great-grandfather Aron Pollak, thanks to his charitable actions, was named Chevalier by the Emperor Franz Joseph, and was granted the name von Rudin (1869). Shortly after the Anschluss in 1938, the family fled from Vienna, suffering under the anti-Jewish oppression of the Nazi regime. Rudin obtained his doctorate in 1947 at Duke University (Durham, North Carolina), and then joined the University

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.4 Synthesis of Causal Filters

163

of Wisconsin, where he moved into a house designed by the celebrated architect Frank Lloyd Wright. In mathematics, Rudin’s name is associated with his description of the closed ideals of the disk algebra, the characterization of the sets of zeros of different Hardy classes in a polydisk (and a ball), the Rudin–Shapiro polynomials, etc. Theorem 5.4.7 (Rudin, 1956; Carleson, 1956) Let σ = σ ⊂ T, m(σ) = 0. For every function f ∈ C(σ) there exists ϕ ∈ Ca (D) such that ϕ | σ = f and ϕCa (D) =  f C(σ) . Proof Let R f = f |σ be the restriction operator, R : Ca (D) → C(σ). By Banach’s theorem (see Appendix E) RCa (D) = C(σ) if and only if there exists c > 0 such that R∗ μ(Ca )∗ ≥ cμ(C(σ))∗ for all μ ∈ (C(σ))∗ . Moreover, Banach’s theorem states that, in the case where the condition is satisfied, for every f ∈ C(σ) there exists a solution of the equation Rϕ = f , ϕ ∈ Ca (D) such that ϕ ≤ c−1  f . The dual space (C(σ))∗ is the space of complex measures M(σ) equipped with the variation norm, μ(C(σ))∗ = Var(μ) (Riesz representation theorem, Appendices A and D). The dual of a subspace Ca ⊂ C(T) is the quotient space of the dual (C(T))∗ , (Ca )∗ = (C(T))∗ /(Ca )⊥ where the annihilator (Ca )⊥ is defined by (Ca )⊥ = {μ ∈ (C(T))∗ = M(T) : 0 = zn , μ = ! μ(−n) every n ≥ 0}. By the Riesz brothers’ Theorem 1.5.4, (Ca )⊥ = H01 = {h ∈ H 1 : h(0) = 0}, hence (Ca )∗ = M(T)/H01 . It is easy to see that the adjoint of a restriction is an embedding operator: R∗ : M(σ) → M(T)/H01 , μ −→ μ + H01 (μ ∈ M(σ)). Consequently, R∗ μ = inf μ + h · mM , h∈H01

where  · M = Var. However, as m(σ) = 0, the measure μ is singular, thus μ + h · mM = μM + h · mM , hence R∗ μ = μM . Therefore, the Banach condition is satisfied with c = 1, which concludes the proof. 

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

164

Harmonic Analysis and Stationary Filtering

5.4.3 Helson Sets: Arbitrary Frequency Response on σ ⊂ T We treat here the case of stable filters and provide an example of a Helson set σ (allowing arbitrary responses on the frequencies of σ) and an example of a non-Helson set. For this purpose, we again use the same theorem of Banach (see the proof of Theorem 5.4.7): if R f = f | σ is the restriction operator, R : Wa → C(σ), then RWa = C(σ) if and only if R∗ μ(Wa )∗ ≥ cμ(C(σ))∗ . Since Wa = F −1 l1 (Z+ ), the dual (Wa )∗ = (F −1 l1 (Z+ ))∗ is realized as l∞ (Z+ ) with the duality expressed as

!  f, c = f (k)ck , c = (ck )k≥0 ∈ l∞ (Z+ ). k≥0

Hence, a set σ is Helson if and only if there exists a constant c > 0 such that, for any μ ∈ M(σ) = (C(σ))∗ , we have sup |! μ(k)| ≥ cμM(σ) . k≥0

Another definition: a set σ ⊂ T is said to be independent (on the field Q) if, for every λ j ∈ σ, λ j  λk , the equation λn11 λn22 . . . λns s = 1, where n j ∈ Z, implies n1 = n2 = · · · = n s = 0. (1) (Helson, 1954). A closed and at most countable independent set σ ⊂ T is a Helson set, i.e. C(σ) = Wa | σ. Proof We write the set σ as a sequence, σ = {λ j : j = 1, 2, . . .}, and let μ ∈ M(σ) and j ∈ T be such that

j μ({λ j }) = |μ({λ j })|. We then make use of the following approximation theorem, known as the Kronecker “Solenoid Theorem”. Theorem 5.4.8 (Kronecker “Solenoid Theorem”) If ζ = (ζ1 , ζ2 , . . . , ζn ) ∈ Tn where {ζ1 , ζ2 , . . . , ζn } is independent, then the trajectory (the semigroup) {ζ k = (ζ1k , ζ2k , . . . , ζnk ) : k = 0, 1, 2, . . .} is dense in Tn . We will use this theorem, but for the proof we refer the reader to Kahane and Salem (1963, pp. 21, 175). We apply Kronecker’s theorem to ζ = (λ1 , . . . , λn ). Then there exists k = kn such that |λkj − j | < 1/n for 1 ≤ j ≤ n, which implies limn λkjn = j for every j.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.4 Synthesis of Causal Filters

165

By the dominated convergence theorem (Appendix A)   sup |! μ(k)| ≥ lim |! μ(k)| ≥ lim  λkn dμ(λ) n→∞ T k→∞ k≥0



j μ({λ j }) = |μ({λ j })| = μM(σ) . = j≥1

j≥1

By Banach’s theorem cited above, we obtain that for every f ∈ C(σ) there  exists ϕ ∈ Wa such that ϕ | σ = f and ϕWa =  f C(σ) . (2) (Helson, 1954). Let σ ⊂ T be a closed set containing arbitrarily long arithmetic progressions, i.e. for every n ≥ 1 there exist ζ, λ ∈ T such that ζλ j ∈ σ and λ j+1  1 for 0 ≤ j ≤ n. Then σ is not a Helson set (C(σ)  Wa | σ). Proof We begin by using Exercise 5.6.2(c): there exist polynomials pn ∈ Ca (D) such that deg(pn ) = n, pn ∞ = 1, limn pn Wa = ∞. Let pn (z) =

n

a jz j

j=0

(where of course, a j = a j,n ). We then consider the points (which exist by hypothesis) ζλ j ∈ σ for 1 ≤ j ≤ n (λ = λn , ζ = ζn ). Set μn =

n

a j δζλ j .

j=0

Then, μn ∈ M(σ) and ! μn (k) =

n

a j (ζλ j )k =

j=0

n

a j ζ k (λk ) j = ζ k pn (λk ).

j=0

 Thus supk∈Z |! μn (k)| ≤ 1; however, μn M(σ) = nj=0 |a j |, hence limn μn M(σ) =  ∞. By Banach’s theorem cited above, we obtain C(σ)  Wa |σ. (3) Examples. It is easy to find sets, countable or not, satisfying either the hypothesis of (1), or that of (2). In general, it is very easy to choose an independent sequence σ = (λ j ) of behavior on T prescribed in advance by using the following recursive construction. λ1 ∈ T \ {1} is arbitrary; if σn = {λ1 , . . . , λn } is already chosen, we consider the subgroup of T generated by σn , Gn = {λk11 . . . λknn : k j ∈ Z} (Gn is a countable set) and select λn+1 ∈ T \ Gn (in particular, this last set is dense everywhere, thus there is much freedom in the choice of λn+1 ). The result σ = (λ j ) j≥1 is an independent sequence. Here are some concrete examples.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

166

Harmonic Analysis and Stationary Filtering

Example of an independent sequence. Let λk = eiθk where θk = πk+1 , k = 1, 2, . . . . Then (λk )k≥1 is an independent sequence. s Indeed, if λn11 λn22 . . . λns s = 1, where n j ∈ Z, then k=1 nk πk+1 = 2πN where N ∈ Z. Then all the nk are zero, since π is not the root of any polynomial with rational coefficients (π is a transcendental number). Example of a convergent sequence containing arbitrarily long arithmetic progressions. Let λk = eiθk , where for 2n ≤ k < 2n+1 θk =

k − 2n 1 − n , n 2 n(n + 1)

n = 1, 2, . . . .

Clearly, limk λk = 1, the sequence θk is monotonically decreasing to 0, and for 0 ≤ j < 2n , we have λ2n + j = ζn tnj where ζn = ei/n , tn = exp(−i/2n n(n + 1)).

5.4.4 Causal Recursive Filters A filter Φ is said to be recursive if the input and output signals Φx = y satisfy a recurrence equation: m m



b j yk− j = a j xk− j , j=0

j=0

where k ∈ Z and a j , b j ∈ C. By using the same techniques of Fourier transforms, it is easy to treat this special case (which appears quite often in practical applications).   Theorem 5.4.9 Let p(z) = mj=0 a j z j , q(z) = mj=0 b j z j , and let Φ be a filter satisfying a recursion as above. Let k p , kq be the zero divisors of p, q (defined in Remark 2.4.4). (1) Φ is a stationary filter if and only if, for every ζ ∈ T, k p (ζ) ≥ kq (ζ). (2) If the above condition is satisfied, Φ is stable and with finite power, and its transfer function is ϕ = p/q. (3) Φ is causal if and only if k p (ζ) ≥ kq (ζ) for ζ ∈ D; Φ realizes a minimum for the weighted phase variation P(ϕ) (of Theorem 5.4.5) if and only if k p (ζ) = kq (ζ) for ζ ∈ D and k p (ζ) ≥ kq (ζ) for ζ ∈ T. Proof By passing to the frequency domain, the recursion becomes qF −1 y = pF −1 x. The rest follows from Theorems 5.2.1, 5.2.2, 5.4.5, and Lemma 5.4.1. The stability is a consequence of the fact that a rational function p/q bounded  on T is in C ∞ (T) ⊂ W.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.5 Inverse Problem: “Can One Hear the Shape of a Drum?”

167

5.5 Inverse Problem: “Can One Hear the Shape of a Drum?” The title of the famous article of Mark Kac “Can one hear the shape of a drum?” (Kac, 1966) explained the spirit of inverse problems to the general public. For filtering, the problem is posed in the following manner. We are confronted with an unknown filter Φ (a “black box,” for example, the atmosphere of a distant planet, an optical filter, etc.), to be identified with the aid of an input test signal x whose output signal y can be observed (or possibly, only a portion of y, or certain functions of y). The objective is to choose a test signal x so that such an experiment leads to the recognition of the transfer function ϕ. To be more precise, here are different situations in which we would like to recognize an unknown filter: (1) (2) (3) (4)

among all the stationary filters, among the physically realizable (causal) filters, when knowing the whole output y = (yn )n∈Z , when knowing only the “physically observable” part y+ = (yn )n≥0 of the output, (5) when knowing a moving average of y.

We begin with the situation where only the energy densities |F −1 x|2 , |F −1 y|2 of the signals x, y are known. Theorem 5.5.1 (identification of a finite-power causal filter from its energy spectrum) Let x, y ∈ l2 (Z) and let Φ be an unknown finite-power causal filter such that Φx = y. (1) For Φ to be uniquely determined by the knowledge of the energy spectra |F −1 x|2 , |F −1 y|2 , it is necessary that |F −1 x(ζ)|2  0 a.e. on T. (2) Suppose |F −1 x(ζ)|2  0 a.e. on T. For Φ to be uniquely determined (up to a multiplicative constant) by the knowledge of |F −1 x|2 , |F −1 y|2 it is necessary and sufficient that Φ be a filter with minimal weighted phase variation (as in Theorem 5.4.5). Proof (1) Let ϕ be a transfer function of Φ. Then, |ϕ| · |F −1 x| = |F −1 y|. If measure, then |ϕ| remains arbitrary |F −1 x(ζ)|2 = 0 on a set E of positive  on E (but nonetheless satisfying E log |ϕ| dm > −∞), and hence ϕ is not well-defined by the equation |ϕ| · |F −1 x| = |F −1 y|. (2) Given Theorem 5.4.5, this part simply affirms that a function ϕ of H ∞ is uniquely determined by its modulus if and only if it is outer. 

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

168

Harmonic Analysis and Stationary Filtering

Theorem 5.5.2 (identification of a finite-power filter by its response to a test signal) Let x ∈ l2 (Z). I. Identification by the complete response. The following assertions are equivalent. (1) Φx = Ψx ⇒ Φ = Ψ for any finite-power filters Φ, Ψ. (2) F −1 x  0 a.e. on T. II. Identification by the physically observable response. The following assertions are equivalent. (1) (Φx)n = (Ψx)n , ∀n ≥ 0 ⇒ Φ = Ψ for any finite-power filters Φ, Ψ. (2) The energy density does not vanish, F −1 x  0 a.e. on T, but the “signal entropy” is infinite:  log |F −1 x| dm = −∞. T

Proof I. Let ϕ, ψ be the transfer functions of Φ and Ψ, respectively (arbitrary functions of L∞ (T)). Then the equation is ϕF −1 x = ψF −1 x, and this implies ϕ = ψ a.e. if and only if F −1 x  0 a.e. on T. II. Here, the equation is P+ ϕF −1 x = P+ ψF −1 x, hence P+ ((ϕ − ψ)F −1 x) = 0 where P+ is the Riesz projection (the orthogonal projection on H 2 ). This is equivalent to g := (ϕ − ψ)F −1 x ∈ H−2 . Let G ∈ H 2 be such that G = g. (2) ⇒ (1) Indeed, (2) implies   log |G| dm = log |(ϕ − ψ)F −1 x| dm T T  log |F −1 x| dm = −∞, ≤ log ϕ − ψ∞ + T

hence G = 0 (see Corollary 2.3.3), and then ϕ = ψ. (1) ⇒ (2) If F −1 x = 0 on a set of positive measure, then the identification is impossible, by part I. Suppose, on the contrary,  log |F −1 x| dm > −∞, T

and set h = min(1, |F −1 x|), ϕ = z[h]/F −1 x, where [h] is an outer function of absolute value h. Then, ϕ ∈ L∞ (T) and ϕF −1 x ∈ H−2 . The filter Φ corresponding to ϕ satisfies Φ  0 and (Φx)n = 0 for all n ≥ 0. This is a contradiction. 

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.5 Inverse Problem: “Can One Hear the Shape of a Drum?”

169

Theorem 5.5.3 (identification of a finite-power causal filter) Let x ∈ l2 (Z). I. Identification by the complete response. The following assertions are equivalent. (1) Φx = Ψx ⇒ Φ = Ψ for any finite-power causal filters Φ, Ψ. (2) x  0. II. Identification by the physically observable response. The following assertions are equivalent. (1) (Φx)n = (Ψx)n , ∀n ≥ 0 ⇒ Φ = Ψ for any finite-power causal filters Φ, Ψ. (2) F −1 x  (H−2 : H ∞ ), where (H−2 : H ∞ ) := {g/h : g ∈ H−2 , h ∈ H ∞ }. Proof I. As in Theorem 5.5.2, the equation is ϕF −1 x = ψF −1 x, but this time ϕ, ψ ∈ H ∞ . Moreover x  0 ⇔ (F −1 x  0 on a set E, mE > 0). Since ϕ = ψ on E, we have ϕ = ψ (see Theorem 1.4.4 or Corollary 2.3.3). II. As in Theorem 5.5.2, the equation is P+ ((ϕ − ψ)F −1 x) = 0, i.e. (ϕ − ψ)F −1 x ∈ H−2 . Hence ϕ − ψ  0 implies F −1 x = h/(ϕ − ψ), where h ∈ H−2 , thus F −1 x ∈ (H−2 : H ∞ ). Conversely, if F −1 x = h/ϕ where h ∈ H−2 , ϕ ∈ H ∞ (ϕ  0), then ϕF −1 x ∈ H−2 , and hence there exists a causal filter Φ such that (Φx)n = 0 for n ≥ 0, however Φ  0. 

5.5.1 Moving Averages of a Signal Let x, y ∈ l2 (Z), where x is interpreted as a finite energy signal and y as measurement instrument making observations at times n ∈ Z. The results of the observation are written (τn x, y)l2 (Z) , and the scalar product is called a moving average of x. In the following theorem we consider a version of the identification problem with the aid of moving averages corresponding to conditions (1)&(4) in the introduction of § 5.5. For other combinations of conditions, see § 5.6. Theorem 5.5.4 (identification of a finite-power filter by means of the physically observable moving average) Let x, y ∈ l2 (Z). The following assertions are equivalent. (1) (τn Φx, y)l2 (Z) = (τn Ψx, y)l2 (Z) ∀n ≥ 0 ⇒ Φ = Ψ for any finite-power filters Φ, Ψ.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

170

Harmonic Analysis and Stationary Filtering

(2) F −1 x  0 and F −1 y  0 a.e. on T and at least one of the signals has infinite entropy:   log |F −1 x| dm = −∞, or log |F −1 y| dm = −∞. T

T

Proof By linearity, we can always suppose Ψ = 0, hence the equation in question is (τn Φx, y)l2 (Z) = 0 for every n ≥ 0. After Fourier transformation, this is equivalent to (zn ϕ f, g)L2 (T) = 0 for n ≥ 0, where f = F −1 x ∈ L2 (T), hence to

g = F −1 y ∈ L2 (T),

 T

zn ϕ f g dm = 0 (n ≥ 0).

The latter is equivalent to saying that ϕ f g ∈ H−1 . Now, it is easy to show the equivalence of the stated properties. (2) ⇒ (1) First note that for functions f, g ∈ L2 (T), we have      log | f g| dm = −∞ ⇔ log | f | dm = −∞, or log |g| dm = −∞ , T





T

T

since T log | f | dm < +∞, T log |g| dm < +∞. Let ϕ f g ∈ H−1 , where f, g ∈ L2 (T) and ϕ ∈ L∞ (T). We use the same inequality as in Theorem 5.5.2 and the hypothesis of (2):   log |ϕ f g| dm ≤ log ϕ∞ + log | f g| dm = −∞, T

T

hence ϕ f g = 0 (by Corollary 2.3.3) and thus ϕ = 0. (1) ⇒ (2) Suppose the contrary. If f g = 0 on a set E ⊂ T, mE > 0, then we obtain ϕ f g = 0 with ϕ = χE  0, which is a contradiction. Suppose f g  0 a.e., but T log | f g| dm = −∞. As in Theorem 5.5.2, set h = min(1, | f g|) and ϕ = z[h]/ f g. Then, ϕ ∈ L2 (T), ϕ  0 and ϕ f g ∈ H−1 ./ This is a contradiction. 

5.6 Exercises 5.6.1 Identification of Filters: Moving Averages I. Let x ∈ l2 (Z). Describe the moving averages n −→ (τn Φx, y)l2 (Z) , (a) over all time (∀n ∈ Z), or

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.6 Exercises

171

(b) physically observable (n ≥ 0), that can identify an unknown finite-power filter Φ among (1) all filters, (2) the causal filters. II. Let y ∈ l2 (Z) define a moving average n −→ (τn Φx, y)l2 (Z) , (a) over all time (∀n ∈ Z), or (b) physically observable (n ≥ 0). Describe the input signals x ∈ l2 (Z) that can identify with the aid of these averages a finite-power filter among (1) all filters, (2) the causal filters.

5.6.2 The Non-equality C a (D)  W a (D) Recall that Ca (D) = H ∞ ∩ C(T) is the disk algebra and Wa (D) is the analytic Wiener algebra of absolutely convergent Taylor series on T:





! |! f (n)| < ∞ . f (n)zn :  f Wa = Wa (D) = F −1 l1 (Z+ ) = f = n≥0

n≥0

Clearly Wa (D) ⊂ Ca (D) and  f Ca ≤  f Wa for every function f ∈ Ca (D). We propose to work on several different proofs of the fact that Ca (D) is much larger than Wa (D). Even though at first sight the subject appears quite special, the question of differentiating Ca and Wa is basic, and at the time of the birth of modern analysis it played an important role. In particular, it is linked to the question of the convergence of the Fourier series of an arbitrary continuous function. The first counter-example, by Paul du Bois-Reymond in 1873, greatly surprised his contemporaries; we will find it as a corollary of our calculations. (a) Show that Ca (D)  Wa (D) if and only if # "  f Wa : f ∈ Pa , p  0 = ∞. c := sup  f Ca Solution: If c < ∞, the norms  · Wa and  · Ca are equivalent, and as Wa is a complete space and dense in Ca , we have Wa = Ca . Conversely, if Wa = Ca , the

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

172

Harmonic Analysis and Stationary Filtering

embedding j( f ) = f, j : Ca → Wa is closed (as the inverse of a continuous mapping) and hence bounded (closed graph theorem, Appendix E). 

(b) Let n

sin(kx)

S n (eix ) =

k

k=1

=

n

eikx − e−ikx

2ik

k=1

.

Show that S n ∞ ≤

5π + 1 < 5. 4

Hint Use the Abel transformation n

k=1

ak bk =

n−1

(ak − ak+1 )Bk + an Bn

where Bk =

k

b j,

1

k=1

and first show that with b j = sin( jx) we have Bk =

sin(kx/2) · sin((k + 1)x/2) . sin(x/2)

Solution: Clearly, it suffices to bound S n (x) above for 0 < x < π. By setting b j = sin( jx) we obtain Bk sin(x/2) = (1/2)(cos(x/2) − cos((2k + 1)x/2)) = sin(kx/2) · sin((k + 1)x/2). Then Sn =

n−1

k=1

1 1 Bk (x) + Bn , k(k + 1) n

where |Bn | ≤ n and hence

|S n | − 1 ≤

(kx/2)((k + 1)x/2)

1 1 1 + k(k + 1) (x/π) k(k + 1) (x/π) k≤1/x k>1/x 

πx π  1 1 5π π π + − . = ≤ + x= 4 x k k + 1 4 x 4 k≤1/x k>1/x



(c) First proof of Ca (D)  Wa (D). Deduce from (b) and (a) that there exist polynomials pn = zn S n ∈ Ca such that pn ∞ ≤ 5 and pn Wa ≥ log(n + 1) (n = 1, 2, . . . ), and hence Ca (D)  Wa (D). Solution: Indeed, pn Wa = from (a) above.

n 1

k−1 ≥

 n+1 1

x−1 dx = log(n + 1), and the rest follows 

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.6 Exercises

173

(d) Second proof of Ca (D)  Wa (D) (du Bois-Reymond, 1873). Deduce from the solution of (c) that there exists a function f ∈ Ca whose Fourier series diverges at the point 1. Of course, for such an f we have f ∈ Ca \ Wa .  Solution: A functional Qn ( f ) = 0≤k≤n ! f (k) (the partial sum of the Fourier series f (k)| ≤  f ∞ ), and at the point 1) is clearly continuous in Ca (T) (since, for example, | ! hence by the Banach–Steinhaus theorem (Appendix E), % $ % $ ∀ f ∈ Ca (T) sup |Qn ( f )| < ∞ ⇔ sup Qn  < ∞ . n

n

However, by (c) (and with its notation), we have |Qn (pn )(1)| =

n

1 1 ≥ log(n + 1), 2k 2 k=1

and pn ∞ ≤ 5. Hence, by definition of the norm, Qn  ≥ result cited, ∃ f ∈ Ca (T) such that supn |Qn ( f )| = ∞.

1 10

log(n + 1), and by the 

(e) Third proof of Ca (D)  Wa (D) (Hardy and Littlewood, 1916). (1) Show that the series f =

eik ln k zk , α+1/2 k k≥2

where 0 < α < 1, converges uniformly on D and represents a function of Ca (D) satisfying the Lip(α) condition, | f (z) − f (z )| ≤ C|z − z |α ; and hence, for 0 < α ≤ 1/2, we have f ∈ Ca (D) \ Wa (D). (2) Show that the series

z k k , f = k k∈Z\{0} where k = |k|  0, converges uniformly on T and represents a function  of C(T); and hence f ∈ C(T) \ W(T) if k≥1 ( k /k) = ∞. Solution: For the solution of (1) (which demands considerable effort), we refer the reader to the treatise of Zygmund (1959, Ch. 5, § 4), or to Hardy and Littlewood (1916). For (2), by performing an Abel transformation, we obtain n−1

zk k

= ( k − k+1 )S k + n S n , k −n≤k≤n k=1 k0

 where  n S n ∞ ≤ 5 n , whereas the series k≥1 ( k − k+1 )S k converges normally,



( k − k+1 )S k ∞ ≤ 5 ( k − k+1 ) = 5 1 < ∞. k≥1

k≥1

The result follows.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press



174

Harmonic Analysis and Stationary Filtering

(f) The Riesz projection P+ on C(T) and Pn . Show that the Riesz projection  f (k)zk is not bounded on the space of polynomials P equipped P+ f = k≥0 ! with the norm  · ∞ , hence is not bounded on C(T). Moreover, 1 log(n + 1), 10 where Pn is the space of trigonometric polynomials of degree ≤ n equipped with the norm  · ∞ . P+ : Pn → Pn  ≥

Hint This is a slight variation on the theme of (d) above; compare the result and the proof with Exercises 2.8.3(g) and 2.8.4(g). Solution: Indeed, S n ∈ Pn and (P+ S n )(eix ) = thus P+ S n ∞ ≥ |(P+ S n )(1)| =

n

eikx , 2ik k=1

n

1 ≥ (1/2) log(n + 1), 2k k=1



and the result follows.

(g) Fourth proof of Ca (D)  Wa (D). Deduce from (f) and (a) that Ca (D)  Wa (D). Solution: If Ca (D) = Wa (D), the norms  · Wa and  · Ca would be equivalent, as would be  · W and  · C (the translations f −→ zn f are isometric in C and W). As the Riesz projection P+ is clearly bounded, and even contracting, on W, this would have to be the same in C, which is not the case by (f). 

(h) Fifth proof of Ca (D)  Wa (D): Littlewood’s crocodile. The sketch of a crocodile (Littlewood, 1953, § 16, p. 46) represents a Jordan domain CRO ⊂ C having the form of a crocodile, whose nose ends at z = 1, and whose teeth overlap (say, at half their length) and have infinite total length. Let f : D → CRO be a conformal mapping such that f (1) = 1. Show that f ∈ Ca (D)\Wa (D). Solution: Indeed, f exists by a classical theorem of Riemann, and by Carath´eodory’s theorem (Rudin, 1998, Theorem 14.19, p. 336), f ∈ Ca (D). Moreover, the image f ([0, 1)) = γ is a curve joining f (0) ∈ CRO and 1 = f (1). Its length |γ| is infinite because γ goes around the teeth, hence  1 

 1    ! | f (r)| dr = ∞ = |γ| = f (k)krk−1  dr  0

 ≤

0

0

1

k≥0

k≥0

|! f (k)|krk−1 dr =

k≥1

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

|! f (k)|.



5.6 Exercises

175

The Riemann conformal mapping f of the unit disk D on the interior of Littlewood’s crocodile gives an example of a function in the disk algebra whose Fourier series does not converge absolutely.

5.6.3 Helson Sets in the Disk D (Vinogradov, 1965) Let σ ⊂ D. Then Wa (D) | σ ⊂ C(σ) | σ, and Wa (D) | σ = C(σ) | σ if and only if σ is finite. Solution: Indeed, if σ is finite, we clearly have Wa (D)|σ = C(σ) (for example, by Lagrange interpolation ). To show the converse, we begin as in Exercise 5.6.2(a) by remarking that Wa (D)|σ = C(σ)|σ if and only if there exists a constant c that controls the norms of the interpolating functions: ∀ f ∈ C(σ) ∃g ∈ Wa (D) such that g | σ = f |σ and gWa ≤ c f C(σ) . If we suppose σ infinite and a Helson set, (Wa (D) | σ = C(σ) | σ), we can select two disjoint sequences z j ∈ σ and w j ∈ σ converging to the same point λ ∈ σ (λ  z j , w j ) and consider the functions fn ∈ C(σ) such that fn (z j ) = 1 and fn (w j ) = 0 for 1 ≤ j ≤ n and fn (z j ) = fn (w j ) = 0 for j > n. Then there exists a function gn ∈ Wa (D) such that gn | σ = fn | σ and

|! gn (k)| = gn Wa ≤ c. k≥0

By the Montel compactness theorem (see Appendix B), we can choose a subsequence (gni ) converging uniformly on the compact subsets of the disk D to a holomorphic N |! gni (k)| leads to gWa ≤ c < ∞ function g. A passage to the limit in the sums k=1 and, for every j, g(z j ) = 1 and g(w j ) = 0, which is impossible, since g is continuous at the point λ. 

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

176

Harmonic Analysis and Stationary Filtering

5.7 Notes and Remarks The mathematical theory of filtering began with the works of Wiener (1930, 1933, 1949) and also of Whittaker (1915, 1924) and Kotelnikov (1933), and on the engineering side with the works of Kotelnikov (1956) and Shannon (1948). The survey article by Masani (1966) gives a broad and comprehensible survey of Wiener’s impact on the discipline, as well as numerous historical details on the early period of filtering theory in the USA, and in particular the pre-eminent role of the treatise Extrapolation, Interpolation, and Smoothing of Stationary Time Series (Wiener, 1949). (In fact this work already existed in 1941–1942 but was classified “Top Secret” until 1949 because of the war; among the students and researchers of the day it was known as the “yellow peril” due to its difficulty and to the yellow color of its cover.) A large part of the theory of random processes, and of optimal control can profitably be interpreted and used for the theory of filtering. In particular, the works of Kolmogorov cited in this book had a strong resonance with the subject. In general, the same remark holds for much of Fourier analysis. Note also that the “theoretical” and “applied” domains of filtering are very different: in the latter the algorithmic and numerical aspects prevail. See Butzer (1983) for a survey of the applied aspects. For more advanced and modern techniques related to wavelets, see Kahane and Lemari´e-Rieusset (1998). The present introduction to the subject is centered on the use of Hardy space techniques, which was important historically. Without mentioning the “continuous version” of the theory presented in this chapter, i.e. covering signals defined on R (see for example Papoulis (1984)), the principal omission here is the topic of sampling (“frames”), closely linked with the techniques developed in this book. Based on a recent survey presentation (Higgins, 1996; Higgins and Stens, 1999), we outline below a few details on the links between sampling and the analysis presented so far in this book. In fact, the history of the discovery of sampling – a crucial idea today for the transmission of signals (both theoretical and practical) – is quite short, but full of twists and turns worthy of a detective novel. In particular, we still do not know precisely when and by whom the idea originated – even if we separate the mathematical and applied engineering aspects. The foundation of the mathematical theory of sampling (of “frames”) is habitually attributed to the British mathematician E. T. Whittaker due to his pioneering paper “On the functions which are represented by the expansions of the interpolation theory” (Whittaker, 1915).

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.7 Notes and Remarks

177

Edmund Taylor (E. T.) Whittaker (1873–1956) was a British mathematician, as well as an astronomer and historian of the sciences. He was president of the London Mathematical Society (1928– 1929) and the Edinburgh Mathematical Society (1914), Copley Medalist of the Royal Society (the most prestigious scientific distinction of Great Britain), and a Fellow of the Royal Society (1905) and the Pontifical Academy of Sciences (Vatican, 1935). The name Whittaker was taken from a farm in Lancashire where the family had lived since the year 1236. He was knighted in 1945. His mathematical results concern relativity theory, representation theory, special functions, partial differential equations and numerical analysis (in particular, interpolation theory). Whittaker’s most famous article is without doubt his work on sampling cited in this chapter. He is also known for his university textbook A Course of Modern Analysis (1902; from the second edition (1915) co-authored with G. N. Watson) – one of the few mathematics texts in Great Britain to remain in print for 100 years. Another memorable mathematical text is The Calculus of Observations (1924). Whittaker’s name is linked to several objects in representation theory and special functions, such as the Whittaker model, Whittaker functions, and Whittaker integrals. His research students included Hardy, Bateman, Eddington, Littlewood, Watson, and Hodge. Beyond mathematics, Whittaker is celebrated as a historian of science and philosophy. His reference text A History of the Theories of Aether and Electricity (1910, 1954) contains a chapter entitled “The relativity theory of Poincar´e and Lorentz,” which gives a history of relativity paying little attention to the results of Einstein (in particular, the formula E = mc2 is attributed to Poincar´e). Whittaker was a devout Christian who converted to Catholicism in 1930. He gave several public

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

178

Harmonic Analysis and Stationary Filtering

lectures on the relations between science and theology: see Space and Spirit: Theories of the Universe and the Arguments for the Existence of God (1946). However, it was Poussin who, in 1908, discovered the two principal Sampling Theorems A and B (below), not to mention the somewhat less definitive statements of Borel (1898) and Hadamard (1901), which without doubt provided a gateway to the theme of applied analysis (their works contain results equivalent to those of Theorem A). In fact, the roots of the theory are even deeper: we find elements of the sampling theorems in the Poisson summation formula (1820)  

1 ! 2πn f (k) f = a n∈Z a k∈Z (a > 0; | f (x)| ≤ c(1 + x2 )−1 and f  ∈ L1 (R)) and in Cauchy’s trigonometric interpolation formulas (1841). The interested reader can find the exact references in the impressive and informative surveys by Higgins (1985), Butzer, Higgins, and Stens (2000), and Butzer et al. (2011). In the statements below, we follow the terminology of signal processing. Sampling Theorem A If a function f does not contain any frequencies beyond Λ/2 cycles per second (hence if ! f (λ) = 0 for |λ| > Λ/2), it is completely determined by its values on a sequence of instances spaced by 1/Λ seconds: ( f (k/Λ) = 0 ∀k ∈ Z) ⇒ f = 0. Sampling Theorem B A function f whose frequency spectrum is limited to [−πΛ, πΛ], i.e. of the form  f (t) =

πΛ

g(x)eixt dx, −πΛ

is the sum of its sampled “cardinal series” f (t) =

k f sinc(Λt − k), Λ k∈Z

where sinc(t) =

sin(πt) πt

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.7 Notes and Remarks

179

is the “sinus cardinal” function, the Fourier transform of χ = χ(−π,π) , ! χ(t) = (2π)−1



χ(x)e−itx dx

R

(= the “spectrum of χ” in the engineering literature). A “downside” of Whittaker’s article is that its level of rigor is not always sufficient: the spectral nature of the cardinal functions f of Theorem B (for Λ = 1) is not made precise, a sufficient condition for the convergence of the same series was not mentioned (he stated that it was sufficient to assume that the function f is entire and bounded on Z, which is not the case), etc. The mathematical evolution of the subject was not “linear”: it turns out that some of the fundamental work of the years 1920–1930 was not digested by the community at the time, leading to a slowdown and duplications in research. In fact, Whittaker’s results were cultivated in two different ways in two corners of the world, far apart and isolated from each other. Namely, in Russia, Kotelnikov, a practicing engineer in radio communications, without any knowledge of the results of Whittaker, rediscovered the formula (Kotelnikov, 1933) and proved (without much rigor in his reasoning) the simple convergence of the cardinal series under the Dirichlet condition for the inversion of the Fourier transform (hence, for integrable and piecewise monotone functions, “which is always the case in electrical engineering,” wrote Kotelnikov). The works of Kotelnikov were unknown to the rest of the world until the end of the 1950s.

Vladimir A. Kotelnikov (1908– 2005) was a Russian (Soviet) mathematician and communications engineer, an inventor and promoter of signal transmission by sampling, and a pioneer of Russian cryptography. His father and grandfather were mathematics professors at the University of Kazan (a region of the Volga); the latter served for a time as an assistant of Nikolai Lobachevsky, one of the pioneers of non-Euclidean geometry. Kotelnikov obtained his university

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

180

Harmonic Analysis and Stationary Filtering

degree at the Moscow Power Engineering Institute (MPEI), and then continued in a postdoctoral position. In 1932 he prepared a presentation for a conference devoted to improving the communication system of the Red Army; the conference never took place but a collection of the presentations was published (1933). It was a pioneering publication by Kotelnikov on the techniques of sampling in signal engineering which (with the works of Claude Shannon 15 years later) definitively changed the landscape of signal processing. Kotelnikov attempted to have his article published in a widely circulated journal, Electricity (in Russian), but it was rejected “because the capacity of the journal is already exceeded and the subject is of limited interest.” Today, sampling based on the Whittaker–Ogura– Kotelnikov–Shannon theorem is ubiquitous in the techniques of signal processing. It was around 1939 that the first communication line using the technique of sampling was launched between Moscow and Khabarovsk (6 100 km). During the Second World War, Kotelnikov worked on problems of encoding/decoding and the encryption of telephonic and radio communications, as scientific director of a team of prisoners of the NKVD (predecessor of the KGB), the infamous Marfino “sharashka” near Moscow. This dark chapter of Stalin’s regime was given literary form in the novel The First Circle by Aleksandr Solzhenitsyn (1968). We note in passing that the novel contains, among other things, the very first description of “wavelets” (a rather vague description, since literary, but quite recognizable), a decomposition technique based on scaling for stable signals (multi-resolution analysis); for a remarkable presentation of signal processing via wavelet techniques, see Kahane and Lemari´e-Rieusset (1998). In 1944 Kotelnikov managed to escape from the confined circle of the prison-laboratories of the NKVD (which was extremely difficult) with the aid of Valeria Golubtsova, director of the MPEI (who by a decisive “coincidence” was the partner of Georgy Malenkov, First Secretary of the Central Committee of the Communist Party). In 1947 he submitted his dissertation, and in 1956 published it (in a more elaborate form) as the monograph Теория потенциальной помехоустойчивости (1956) (published in English as The Theory of Optimum Noise Immunity, 1959). During the years 1950–1970 Kotelnikov guided with great success a certain number of major projects such as the observation and cartography of the planets

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.7 Notes and Remarks

181

(in particular, he was the chief editor of Atlas of the Surface of Venus, 1985) and the communications and control part of the Russian space exploration program. He was elected member of the Academy of Sciences of the USSR (1953), and later became its Vice President (1970–1988). He was also the chief editor of a certain number of journals, and awarded many Soviet and Russian honors, as well as the Eduard Rhein Prize (Germany, 1999) and the IEEE’s Alexander Graham Bell Medal (2000).

Letter from the journal Electricity (in Russian) rejecting Kotelnikov’s pioneering paper “because the capacity of the journal is already exceeded and the subject is of limited interest.” The techniques of sampling in signal engineering would soon become ubiquitous throughout the world.

The Marfino “sharashka” was a Soviet Gulag research laboratory near Moscow (25 Botanicheskaya Street, Marfino) where, during and after the Second World War, important work was done in encoding/decoding and the encryption of telephonic and radio communications on behalf of the secret services NKVD/KGB.

The principle of sampling (hence the contents of Theorem A above) was also known (at latest in 1928) by Harry Nyquist, a Norwegian engineer who worked at Bell Telephone Laboratories in the USA. However, on the mathematical side, it was Kinnosuke Ogura, a renowned Japanese professional mathematician, who was the most advanced, as early as 1920 (Ogura, 1920): he had corrected the inaccuracies of Whittaker by giving a counterexample to his conditions and by replacing them (but without a proper proof) with a sufficient condition (too strong to be necessary) for the convergence of the cardinal series ( f must be an entire function of type less than π bounded on R).

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

182

Harmonic Analysis and Stationary Filtering

Kinnosuke Ogura (1885– 1962) was a Japanese mathematician with a very large spectrum of interests: from interpolation theory (including his publications on sampling) and other questions in analysis, up to relativity and differential geometry. He authored around 70 articles in pure and applied mathematics (including 42 between 1911 and 1923, including some of major importance) and published 35 books (!) on subjects varying from infinite series, applied Kinnosuke Ogura. (From Butzer et al. (2011), reprinted by analysis and relativity, through to social problems, permission of Springer Nature.) education, and history. Ogura studied in Tokyo between 1902 and 1916 (first in chemistry, then in mathematics) and obtained his doctorate in 1916 on a subject of mathematical physics. He spent two years in France, in 1920–1921, motivated in particular by the opportunity to study and collaborate with Borel and subsequently Langevin and Hadamard. His common interests with Borel ranged from the cardinal series to the applications of geometry to relativity. He was invited to lecture at the International Congress in Strasbourg (1920). In 1922, together with physicist colleagues, he organized Einstein’s six-week visit to Japan. Ogura was interested in a broad range of topics in mathematics, but also in philosophy, statistics, and social problems (he was influenced by Marxism and by Leo Tolstoy). During his career as a research professor, he taught at the Siomi Research Institute (Osaka) and at the universities of Hiroshima and Osaka, but his work was often interrupted because of health problems. After his retirement in 1943, he presided over the Society of History of Sciences and Mathematics. With all of these activities, Ogura was one of the key figures of Japanese mathematics between the two

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.7 Notes and Remarks

183

World Wars. A distressing fact in Ogura’s biography is that during the Sino-Japanese War of 1937–1945 (which caused the death of more than 20 million people in Asian countries) he took the side of the military power, in particular, engaging in 1940 with the Imperial Rule Assistance Association (a militarist civilian organization) and publishing several articles which mobilized Japanese scientists for the “Greater East Asia War” (an official slogan). When he published a revised version of these articles in 1948, Ogura replaced the words “for the Greater East Asia War” with “for the Democratic Revolution” (see “The Mathematician K. Ogura and the ‘Greater East Asia War’,” by Tetu Makino (2003)). The moment of truth for the mathematical aspect of sampling came with the Paley–Wiener theorem (Paley and Wiener, 1934: F L2 (−a, a) is a space of entire functions of exponential type ≤ a and square integrable on R; see also § 6.3 below) and, for the cardinal series, with the article of Hardy (1941) where the Whittaker–Ogura–Kotelnikov theorem (above all, a dream . . . ) took on its definitive form: {sinc( · − k) : k ∈ Z} is an orthonormal basis of the space F L2 (−π, π), and hence, for every f ∈ F L2 (−a, a), 



f (k) sinc( · − k), | f |2 dt = | f (k)|2 , f = R

k∈Z

k∈Z

and moreover, F L (−π, π) | Z = l (Z) (“free” interpolation on Z). 2

2

Furthermore, in the same article Hardy assigned a name to this space by calling it the Paley–Wiener space (a name which it still holds): g : g ∈ L2 (−πΛ, πΛ)} PWπΛ = {! (a closed subspace of L2 (R)). He also showed that sinc(t − x) is a reproducing kernel for PWπ ,  f (x) sinc(t − x) dx, ∀ f ∈ PWπ . f (t) = R

We can also mention that a cardinal series of Theorem B above is simply a special case of the formula of Lagrangian interpolation,

f (zk )p(z) , f (z) = p (zk )(z − zk ) k where p is an entire holomorphic function and (zk ) its simple zeros, and which can be justified by the residue theorem of complex analysis.

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

184

Harmonic Analysis and Stationary Filtering

In engineering, the principal practical consequence of Sampling Theorems A and B is that for the treatment (transmission, transformation, etc.) of continuous signals f (t), t ∈ R “not containing rapid oscillations,” we can restrict ourselves to a discrete sequence f (k/Λ) without any loss of information contained in the signal (this is the essence of the digitalization of electrical, optical, and any other type of signals). This discovery was a veritable revolution in signal processing and telecommunications engineering. As a consequence, the evolution of the theory and practice of sampling advanced in leaps and bounds in Russia during the years 1935–1945, especially because during the Second World War signal processing was linked with the encryption and coding of communications (see the biography of Kotelnikov on page 179). The true engineering pioneer of sampling, at the center of all these changes, was Vladimir Kotelnikov; in particular, he implemented the ideas of his seminal work, “On the transmission capacity of ‘aether’ and wire in electrocommunications” (Kotelnikov, 1933), by installing and putting into operation the secure high-resolution telephone line between Moscow and Khabarovsk. On the other side of the Atlantic, in the USA, the principle of sampling was discovered by Claude Shannon, an engineer at Bell Telephone Laboratories.

Claude E. Shannon (1916–2001) was an American electrical engineer, cryptographer, and mathematician, and founder of information theory. He was a distant relative of Thomas Edison, both being descended from John Ogden, one of the founding fathers of the American colonies and ancestor to many famous individuals. One of the greatest inventions of the twentieth century is attributed to Shannon: all information and all communications can be coded in the simple language of 0s and 1s. This idea was presented in his Master’s thesis at MIT, “A symbolic analysis of relay and switching circuits” (1937), described by Howard Gardner (Harvard) as “possibly the most important, and also the most famous, master’s thesis of the century.” (To be fair, note that this discovery had also been made in 1935 by a Russian logician and

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

5.7 Notes and Remarks

185

engineer, Victor Shestakov: it figures in his thesis submitted in 1938 at the University of Moscow and published in 1941.) During the Second World War, Shannon worked at Bell Laboratories, investigating cryptography and the automatic guidance of anti-aircraft fire. Once declassified, his works gave rise to his classical post-war publications “A mathematical theory of communication” (1948, where he introduced the word “bit” for the minimal unit of information, as well as his rediscovery of the Whittaker–Ogura–Kotelnikov theorem), and “Communication theory of secrecy systems” (1949). These publications of Shannon revolutionized the theory and practice of telecommunications (by introducing, in particular, error-correcting codes), cryptography, applied probability and statistics, and then the theory of replication of DNA, etc. The media buzz around this “numerical revolution” was so important that Shannon felt obliged to remark that “Information theory has perhaps ballooned to an importance beyond its actual accomplishments.” Later, Shannon worked on a program to play chess (and published “Programming a computer for playing chess” (Shannon, 1950); the first match was played by the Los Alamos MANIAC machine in 1956), and conducted the very first experiments in artificial intelligence (a “mouse” running a labyrinth with elements of self-learning). He retired early, in 1966, but continued as a consultant for Bell Laboratories. At the end of his life he was stricken with Alzheimer’s disease. Numerous witnesses attest to his acute mind, his sense of humor, and his originality. For example, he enjoyed tearing down the corridors of MIT on a unicycle, scaring the living daylights out of his colleagues, or surprising them with his “Ultimate Machine” – a device whose sole function is to switch itself off. The list of honors received by Claude Shannon has about 30 entries, including the US National Medal of Science (1966) and the Kyoto Prize (1985). Several concepts in computer science bear Shannon’s name: Shannon’s theory of information, the capacity of a transmission channel, the Shannon entropy, etc.

His findings were published around 1948–1949 (probably obtained around 1940, but classified “Top Secret” until 1949) and contained results similar to those of Kotelnikov, again with a presentation from an “engineering perspective,” without discussion of the mathematical conditions of validity, or precise details of the classes of signals or the types of convergence of

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

186

Harmonic Analysis and Stationary Filtering

the developments. As a result, in the USA and in Europe, signal processing was known as “Shannon’s theory,” and it was only very slowly that the more complete history of the subject began to be recognized, with the appearance of the names of Kotelnikov around 1959 (and his article, Kotelnikov (1933), in 2001!) and of Ogura, as late as 1992. For more historical and mathematical details we refer the reader to the surveys of Butzer et al. (2000, 2011), as well as to the book by Kahane and Lemari´e-Rieusset (1998). The results attributed to Wiener, Theorems 5.2.1– 5.2.2 and Lemma 5.4.1, can be found, at times without an explicit statement, in Wiener (1933, 1949). In particular, a significant event of the epoch was the characterization of the energy spectrum of a causal filter by the integrability of the logarithm (Corollary 5.4.2). For Theorem 5.4.7 see Rudin (1956) and Carleson (1956). A large number of generalizations are known: see Garnett (1981), Gamelin (1969), and Havin and J¨oricke (1994). For the Helson sets and related problems, see Kahane and Salem (1963), Rudin (1962), and for the presentation of § 5.5, Nikolski (2002). The contents of Exercise 5.6.2 are classical: see Zygmund (1959).

https://doi.org/10.1017/9781316882108.007 Published online by Cambridge University Press

6 The Riemann Hypothesis, Dilations, and H 2 in the Hilbert Multi-disk

Topics. Euler ζ function, integral representations of ζ, the Riemann hypothesis, the H p spaces in the half-plane, the Paley–Wiener theorem, invariant subspaces generated by the ζ function, distance function and zeros of ζ, Beurling’s problem on the completeness of the dilations, the space H 2 in the Hilbert multidisk, dilations of the polynomials. In 1737 Leonhard Euler, at the time professor at Saint Petersburg (and member of the Russian Academy of Sciences), wrote an article entitled Variae observationes circa series infinitas (published in 1744) where he defined a function that is now called the Euler zeta function (or, more frequently, the Riemann zeta function: see § 6.8 for comments),

1 . ζ(s) = ns n≥1 In his 1737 article, Euler was interested in ζ(s) uniquely for the values s ∈ N: s = 1 to give a new proof of the infinitude of prime numbers, and s = 2, 4, . . . to resolve the “Basel problem,” consisting precisely of the calculation of the sums ζ(2), ζ(4), etc. In 1749, he extended the definition to the real values. Later, in a result published in 1761, Euler presented a crucial tool for the study of ζ – the Riemann functional equation of Theorem 6.1.5, which in all justice should be called the Euler–Riemann equation (see the article by Gelfond (1958) on this subject).

187 https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

188

RH, Dilations, and H 2 in the Hilbert Multi-disk

Leonhard Euler (1707–1783) was a Swiss, Russian, and German mathematician, a mathematical genius, instigator of several modern disciplines, founder of topology and graph theory, author of approximately 900 original articles (many – but not all – were published in Switzerland in 73 volumes in quarto), including very important monographs in mechanics, analysis, naval science, celestial mechanics, integral calculus, algebra, etc., not to mention works for the general public such as Letters to a German Princess (three volumes!) explaining the principles of physics, philosophy, and mathematics in a simplified form. Around 30 mathematical objects bear Euler’s name: the base e (Euler number) of the natural logarithms, the Euler angles, the Euler Γ function, the Euler constant, the Euler–Lagrange equation, etc. Introducing the function ζ(s) (see § 6.1), Euler established the product formula of § 6.1.2 and (de facto) the functional equation of Theorem 6.1.5. His results in number theory, as well as the famous correspondence between Euler and Christian Goldbach (between Berlin and Saint Petersburg/Moscow), defined the research direction for additive number theory for centuries. Leonhard Euler was born in 1707 in Basel (Bˆale). Under the influence of a family friend, Johann Bernoulli, he studied science at the University of Basel, and submitted a Master’s thesis in 1723 on the philosophies of Descartes and Newton. In 1726, under Bernoulli’s supervision, he defended his doctoral thesis on the propagation of sound. After losing the competition (!) for a position at the University of Basel, he accepted an offer as professor of Physiology (!) at the Saint Petersburg Academy of Science, which had just opened the previous year. He remained in Russia for 14 years (1727–1741), and returned later (1766–1783) at the personal invitation of the Russian Empress Catherine the Great, after spending the years 1741–1766 in Berlin as head of the Prussian Academy. Euler

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

189

married Katharina Gsell and had 13 children. He wrote (in fact dictated, because he became blind) 800 pages in quarto per year, in addition to participating in 10 to 15 annual conferences. In Russia, his annual salary grew from 200 roubles at the start of his career to 3 000 roubles by the end (a horse – the equivalent of a car today – cost around 5 roubles); this allowed him to support a household of up to 20 people. Euler had an equable character: when in 1767 he lost the use of his second eye, he confided “Henceforth nothing will be able to distract me from mathematics.” The Marquis de Condorcet quoted Euler’s reply to a question of the German Queen Mother during a reception at the Berlin court after his return from Russia: “Why will you not speak to me?” “Madame,” replied Euler, “because I have come from a country where one can be hanged for what one says.” Another anecdote recounts his confrontation with Diderot, who had been invited by Catherine the Great to the court of Saint Petersburg – a confrontation that Euler is said to have begun by announcing “eiπ = −1, hence God exists: reply!”1 In his book A Concise History of Mathematics, Dirk Struik describes the contemporary response to the publication of Euler’s mathematical theory of music (Tentamen novae theoriae musicae, 1739): “it was too musical for mathematicians and too mathematical for musicians.” Euler’s extraordinary importance was largely recognized, from the time of Laplace, who stated Lisez Euler, lisez Euler, c’est notre maˆıtre a` tous (“Read Euler, read Euler, he is the master of us all”), up to a recent text on the Internet, “Top 10 Greatest Mathematicians” (M. Sexton): “If Gauss is the Prince [of Mathematics], Euler is the King.” A crater on the moon and an asteroid are named after Euler, as well as several scientific prizes and research institutes; many stamps and coins bear his effigy. Euler is buried in the Alexander Nevsky Cemetery in Saint Petersburg. Euler had already linked the function ζ to the principal question of arithmetic: “How many prime numbers are found in nature and how are they distributed?” Over time, this link has only been strengthened. In 1859, Bernhard 1

The anecdote was recounted by Dieudonn´e Thi´ebault in his Souvenirs de vingt ans de s´ejour a` Berlin (vol. 3, p. 142 of the 1804 edition). He does not cite Euler by name and gives for the formula (a + bn )/z = x. He points out he cannot ensure the veracity of this story and he simply transcribed what he heard. If this story has an element of truth, we could suppose that Euler invoked a real formula, illustrating the “divine beauty” of mathematics: this is why it is often replaced (as here) with eiπ = −1.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

190

RH, Dilations, and H 2 in the Hilbert Multi-disk

Riemann directly connected the question of the distribution of prime numbers with the properties of ζ. In particular, he showed that ζ(s) = 1/(s − 1) + F(s), where F is an entire holomorphic function in C (see Theorem 6.1.5 below), and presented the hypothesis: All zeros of ζ in the half-plane Re(s) > 0 are located on the line Re(s) = 1/2. Since then, this proposition has been known as the Riemann hypothesis (RH); it remains unresolved today. In this chapter we present an approach to the RH using the invariant subspaces of the space H 2 discovered by the Swedish mathematician Bertil Nyman, in 1949. This approach is based on an integral representation of the function ζ and the Fourier transform. In what follows, we let ρ(x) denote the fractional part of x ∈ R, ρ(x) = x − [x]

([x] the integer part of x).

The function ρ is 1-periodic, ρ(x) = ρ(x + 1); for x > 0 we set ϕ(x) = ρ(1/x).

6.1 The Euler ζ Function and the Riemann Hypothesis (RH) We outline here a few elementary properties of the Euler ζ function. Definition 6.1.1 Let s ∈ C, Re(s) > 1. Set ζ(s) =

1 ; ns n≥1

this is called the Euler ζ function. Clearly the series converges absolutely (and uniformly on every half-plane Re(s) ≥ 1 + , > 0), and represents a holomorphic function in Re(s) > 1.

6.1.1 Prime Number Decomposition (Euclid, c. 300 BCE; Gauss, 1801) Let (p s ) s≥1 be the sequence of consecutive prime numbers p1 = 2,

p2 = 3,

p3 = 5,

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

...

6.1 The Euler ζ Function and the Riemann Hypothesis (RH)

191

Then, every natural number n ∈ N possesses a unique representation of the form where α j ∈ Z+ = N



n = pα1 1 pα2 2 . . . pαk k , {0} (and k = k(n)).

6.1.2 The Euler Infinite Product For every s ∈ C, Re(s) > 1, ζ(s) =

 k≥1

1−

1 −1 , pks

and the product converges absolutely. Proof Indeed, the product converges absolutely since pk ≥ k and hence  s k≥1 1/|pk | < ∞. Let D(k) be the set of integers having all their prime divisors among p1 , . . . , pk , i.e. D(k) = {pα1 1 pα2 2 . . . pαk k : α j ∈ Z+ , 1 ≤ j ≤ k}. Given the absolute convergence of Definition 6.1.1, we have

1 ζ(s) = lim , k ns n∈D(k)

1



1 = sα1 sα2 s n p p2 . . . pksαk α ≥0...α ≥0 α ≥0 1 n∈D(k) 2

k



=

1



α2 ≥0...αk ≥0

 1 1 −1 1 −1 = , 1 − p1s p sj p2sα2 . . . pksαk j=1 k

1−

and the result follows.



Corollary 6.1.2 ζ(s)  0 for every s ∈ C, Re(s) > 1. This is clear by § 6.1.2 and the definition of a convergent product. Lemma 6.1.3 For every t ≥ 1, Re(s) > 0, we have  1 1 ζ(s) − s = ϕ(tx)x s−1 dx. t(s − 1) t s 0 Proof With u = 1/tx we have  1  ∞  1  ∞  s dx −s −s du −s =t =t ϕ(tx)x ρ(u)u + x u 0 1/t 1/t 1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press



192

RH, Dilations, and H 2 in the Hilbert Multi-disk

(ρ(u) = u for 0 ≤ u < 1) = t−s



1 1/t

= t−s

du

+ u s n≥1



n+1

ρ(u)

n

du  u s+1

  1  du  1  n+1 (u − n) s+1 1 − 1−s + 1−s t u n≥1 n

(integration by parts)   1  1  1  1  n+1 du − 1 − 1−s + 1−s s n u s (n + 1) s t n≥1    1  1  1 ∞ du 1 (ζ(s) − 1) = t−s − 1 − 1−s + 1−s s 1 us s t   1   1 1 1 = t−s − (ζ(s) − 1) 1 − 1−s + 1−s s(s − 1) s t  ζ(s)  1 = t−s − − . s (1 − s)t1−s = t−s



Corollary 6.1.4 The function ζ can be extended to a meromorphic function in the half-plane {Re(s) > −1} having a single pole at the point s = 1 (of residue 1) and the integral representations  ∞ 1 ζ(s) = − ρ(y)y−s−1 dy, for Re(s) > 0, s s−1 1  ∞ ζ(s) 1 =− ρ(y) − y−s−1 dy, for − 1 < Re(s) < 0, s 2 0 a ∞ where the last integral is considered as improper: 0 = lima→∞ 0 . Indeed, since the function ϕ is bounded, the integral 

1

s −→ 0

 ϕ(x)x s−1 dx =

∞

ρ(y)y−s−1 dy

1

is holomorphic for s ∈ C+ (Appendix A). Moreover, for the same s, we have  ∞  ∞  1 ∞ −s−1 1 ρ(y)y−s−1 dy = y dy ρ(y) − y−s−1 dy + 2 2 1 1 1  ∞ 1 1 = ρ(y) − y−s−1 dy + . 2 2s 1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.1 The Euler ζ Function and the Riemann Hypothesis (RH)

193

The last improper integral 

∞

s −→

ρ(y) −

1

1  −s−1 dy y 2

converges uniformly in s on the compact sets in Re(s) > −1, and hence can be extended analytically in the named half-plane: indeed, the function  x 1 R(x) = ρ(y) − dy 2 1  k+1 is bounded for x > 1, since k (ρ(y) − 1/2) dy = 0 for every k and since for any a > 0, an integration by parts leads to  a  a ρ(y) − (1/2) R(y) −s−1 dy = R(a)a + (s + 1) dy. s+1 s+2 y 1 1 y Consequently, after such an extension, for −1 < Re(s) < 0 we obtain   ∞ ζ(s) 1 1 1 −s−1 = − dy − ρ(y) − y s s−1 2 2s 1    ∞  1 1 1 1 −s−1 1 −s−1 − = dy + dy − ρ(y) − ρ(y) − y y s−1 2 2 2s 0 0   ∞ 1 −s−1 dy, =− ρ(y) − y 2 0 where we used the fact that   1  1  1 1 1 −s−1 1 1 −s−1 + . dy = y−s dy − y dy = ρ(y) − y 2 2 0 1 − s 2s 0 0



Theorem 6.1.5 (Euler, 1761; Riemann, 1859) The function s −→ ζ(s) can be extended analytically in the entire plane C, with the exception of a simple pole at s = 1, where it satisfies the Euler–Riemann functional equation ζ(s) =

1 πs (2π) s sin Γ(1 − s)ζ(1 − s), π 2

or ξ(s) = ξ(1 − s), where Γ is the Euler Γ function and ξ(s) =

1 s(s − 1)π−s/2 Γ(s/2)ζ(s). 2

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

194

Proof (Hardy, 1922; Titchmarsh, 1951) By Exercise 2.8.3(b), the Fourier series of  x  1 x−π − = ρ 2π 2 2π is −

sin(kx) k≥1

and thus ρ(y) −

kπ

,

sin(2πky) 1 =− , 2 πk k≥1

y ∈ R \ Z.

The series converges everywhere, and its partial sums are uniformly bounded (see Exercise 5.6.2(b)). Let a > 0 and s ∈ C, −1 < Re(s) < 0. By Corollary 6.1.4,   a 

sin(2πky) −s−1 ζ(s) = lim dy y a→∞ 0 s πk k≥1

 a sin(2πky) = lim y−s−1 dy, a→∞ πk 0 k≥1 and hence the last limit would be equal to

 ∞ sin(2πky) y−s−1 dy πk k≥1 0 (the integrals taken as improper) if we could show that

 ∞ sin(2πky) y−s−1 dy = 0. lim a→∞ πk a k≥1 However, 

∞

a

thus

  

a

sin(2πky) cos(2πka) s + 1 dy = − − 2πk y s+1 2πka1+s

∞

 a

∞

cos(2πky) dy, y s+2

 sin(2πky)  |s + 1| 1 dy ≤ + , y s+1 2πkaRe(s)+1 2πk(Re(s) + 1)aRe(s)+1

hence the required convergence. We thus obtain   ζ(s) 1 1 ∞ sin(2πky) 1 (2πk) s ∞ sin(x) = dy = dx. s π k≥1 k 0 π k≥1 k y s+1 x s+1 0

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.1 The Euler ζ Function and the Riemann Hypothesis (RH)

195

The integrals under the summation sign are known and can be calculated with the aid of the Euler Γ function whose definition for Re(z) > 0 is  ∞ Γ(z) = e−t tz−1 dt. 0

It is well known (see for example Titchmarsh (1939, § 3.1.2.7, § 4.4.1)) that Γ can be extended to a meromorphic function in C satisfying the following equations (among others): π , Γ(z)Γ(1 − z) = sin(πz)   1 1 Γ(2z) = √ 22z−1 Γ(z)Γ z + , 2 π

Γ(z + 1) = zΓ(z),



∞

tz−1 sin(t) dt = Γ(z) sin(zπ/2) (0 < Re(z) < 1).

0

By using these equations, we come to the first formula stated, ζ(s) =

1 s(2π) s Γ(−s) sin(−πs/2) π k1−s k≥1

s(2π) s Γ(1 − s) sin(−πs/2)ζ(1 − s) π −s (2π) s Γ(1 − s) sin(πs/2)ζ(1 − s), = π

=

at least for the values of s in −1 < Re(s) < 0. This formula shows that ζ can be extended analytically in the half-plane Re(s) < 0 while continuing to satisfy the same equation. By once again using the above identities for Γ it is easy to verify that ξ(s) = ξ(1 − s) is an equation equivalent to the preceding one.  Corollary 6.1.6 ζ has zeros ζ(−2n) = 0 (n = 1, 2, . . . ); all the other zeros of ζ (if they exist) are in {s ∈ C : 0 ≤ Re(s) ≤ 1} and are symmetric with respect to the line Re(s) = 1/2. Indeed, this is clear by the equation of Theorem 6.1.5, Corollary 6.1.2 and  the fact that for any z, Γ(z)  0. The −2n, n = 1, 2, . . . , are called the trivial zeros of ζ.

6.1.3 The Riemann Hypothesis (RH), 1859 All the non-trivial zeros of ζ are situated on the line Re(s) = 1/2.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

196

RH, Dilations, and H 2 in the Hilbert Multi-disk

Bernhard Riemann (1826– 1866) was a German mathematician, the most influential creative genius in the mathematical renaissance of the second half of the nineteenth century. The ideas of Riemann definitively transformed complex analysis, geometry, and number theory, and also provided a strong impetus for real harmonic analysis. Three of Riemann’s four most influential works were “qualifying texts”: his doctoral thesis (G¨ottingen, 1851, under the supervision of Gauss) containing the theory of Riemann surfaces and conformal mappings, his habilitation thesis (1853), devoted to Fourier series (with the Riemann integral as a tool), and his famous Habilitationsvortrag (1854, an inaugural habilitation conference) entitled ¨ Uber die Hypothesen, welche der Geometrie zu Grunde liegen (chosen by Gauss from the three themes proposed by Riemann). These three ¨ masterpieces were published posthumously. The fourth work was “Uber die Anzahl der Primzahlen unter einer gegebenen Gr¨osse” (1859) where Riemann introduced his ideas on the role of the ζ function in the complex plane in the distribution of prime numbers (this was his only opus devoted to number theory). These contributions of Riemann became – and remain – absolutely fundamental for the mathematics and physics of the nineteenth to twenty-first centuries. An astronomical number of publications are devoted to the development of Riemann’s ideas and results. For a presentation intended for the general public, see for example Bernhard Riemann 1826–1866: Turning Points in the Conception of Mathematics by Detlef Laugwitz (Birkh¨auser, 2008), Riemann, Le g´eom`etre de la nature by Rossana Tazzioli (vol. 12 of Pour la Science, 2002), or “Riemann” by Hans Freudenthal (in Dictionary of Scientific Biography, 2008). As remarked in the last of these, “Riemann’s evolution was slow and his life short.” He only managed to write around 15 mathematical manuscripts, but these rare works opened a new era in mathematics. Riemann’s name is associated with dozens of

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.2 An Approximation Implying the Riemann Hypothesis

197

important concepts: Riemannian geometry, Cauchy–Riemann equations, Riemann surfaces, the Riemannintegral,Riemannconformal mapping theorem, Riemann– Hilbert problem and method, Riemann hypothesis, Riemann– Lebesgue lemma, Riemann sphere, etc. In particular, Riemannian geometry was fundamental to the creation of general relativity – and also in the inspiration of a certain mathematician Charles Dodgson (better known under his literary pseudonym Lewis Carroll) for his ingenious Alice’s Adventures in Wonderland (1865) and Through Riemann’s seminal contributions to the Looking-Glass (1871). geometry likely inspired Lewis Riemann’s career was slow and Carroll, otherwise known as Oxford brief: he became a professor at mathematics lecturer Charles G¨ottingen only in 1859 (after Dodgson, when he wrote Alice’s the death of Dirichlet), and alAdventures in Wonderland and ways suffered from a lack of stuThrough the Looking Glass. Carroll’s dents (his renowned course on self-caricature is (presumably) Abelian functions was frequented entitled “Me when I am lecturing” – by only three students, includperhaps on Riemann’s curved ing Dedekind). A deterioration spaces and imaginary numbers. in his health (latent tuberculosis?) frequently forced him to seek refuge in Italy (1862–1866). Riemann was married in 1862 to Elise Koch, with whom he had a daughter.

6.2 An Approximation Implying the Riemann Hypothesis In what follows, the following notation is used:  V = Lin(ϕ(tx) : t > 1), V0 = f ∈ V : f (1) = 0 ,

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

198

where, recall, ϕ(x) = ρ(1/x) = 1/x − [1/x] (x > 0). In the following theorem, and for the remainder of this chapter, we consider a space L2 (E), where E ⊂ R, as the subspace of L2 (R) consisting of the functions of L2 (R) that are zero on the complement R \ E. Theorem 6.2.1 Let d = distL2 (0,1) (1, V0 ). (1) The disk

 Dd = s ∈ C : d2 |s|2 < 2 Re(s) − 1

(⇔ (x − 1/d2 )2 + y2 < r2 := 1/d2 (1/d2 − 1) where s = x + iy) does not contain any zeros of s −→ ζ(s). If d > 0, then Dd = D(1/d2 , r). (2) (Nyman, 1950) d = 0 ⇔ 1 ∈ closL2 (0,1) (V0 ) ⇒ (RH) (i.e. all the zeros of ζ in Re(s) > 0 are on the line Re(s) = 1/2). Proof (1) Suppose s ∈ C, Re(s) > 0, ζ(s) = 0   and f ∈ V0 , f (x) = nk=1 ak ϕ(tk x), where tk ≥ 1 and 0 = f (1) = nk=1 ak /tk . Then, by Lemma 6.1.3,  1  1 n 1

(1 − f (x))x s−1 dx = − ak ϕ(tk x)x s−1 dx s k=1 0 0 =

n n

1 1

1 − ζ(s) ak /tks − ak /tk = , s s − 1 k=1 s k=1

and hence  1/2  1    1  1  =  (1 − f (x))x s−1 dx ≤ 1 − f  2 2s−2 |x |dx  L (0,1)  s   0 0 = 1 − f L2 (0,1)

1 . (2 Re(s) − 1)1/2

Passing to the infimum over f , we obtain 2 Re(s) − 1 ≤ d2 |s|2 , i.e. s  Dd . (2) If d = 0, this is the half-plane Dd = {s ∈ C : Re(s) > 1/2}, which is free of zeros. The RH follows by the symmetry in Corollary 6.1.6.  Remark 6.2.2 Clearly, with an arbitrary function f ∈ V0 , we obtain a disk Dd , d ≤ d = 1 − f L2 (0,1) free of zeros of ζ. For d small enough, the most left-hand point of Dd has for abscissa xd =

1 1 d2 + + o(d2 ), − r = 2 8 d2

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.3 H 2 (C+ ) and the “Weak Paley–Wiener Theorem”

199

√ and the radius r = ( 1 − d2 )/d2 tends to ∞, hence the disks Dd fill the halfplane {z ∈ C : Re(z) > 1/2}. We conclude this section with an interpretation of the approximation of Theorem 6.2.1 in terms of invariant subspaces of the semigroup of dilations. Lemma 6.2.3 With the notation of Theorem 6.2.1, the following assertions are equivalent. (1) (2) (3) (4) (5)

d = 0. 1 ∈ closL2 (0,1) (V0 ). χ(0,1) ∈ closL2 (0,∞) (V0 ). χ(0,1) ∈ closL2 (0,∞) (V). closL2 (0,∞) (V0 ) = L2 (0, 1).

Proof The equivalence (1) ⇔ (2) is evident. Since for f ∈ V and x > 1 we have f (x) = f (1)/x, we thus obtain the equivalences (2) ⇔ (3) ⇔ (4) (we consider L2 (0, 1) as a subspace of L2 (0, ∞) consisting of functions f ∈ L2 (0, ∞) that are zero on (1, ∞)). Evidently, (5) ⇒ (3). For the converse, (3) ⇒ (5), observe that V0 is a vector subspace of L2 (0, 1) invariant by dilatation Dt , t ≥ 1, Dt f (x) = f (tx)

(x > 0, t ≥ 1),

Dt V0 ⊂ V0 , t ≥ 1. Indeed, if f ∈ V, f (1) = 0 then Dt f ∈ V and Dt f (1) = f (1)/t, hence Dt f ∈ V0 . Consequently, a subspace E = closL2 (0,∞) (V0 ) is also Dt -invariant, and by hypothesis χ(0,1) ∈ E. Then, for every t ≥ 1, Dt χ(0,1) = χ(0,1/t) ∈ E, and hence every step function on the interval (0, 1) is in E. However, the space of step  functions is dense in L2 (0, 1), hence E = L2 (0, 1).

6.3 H2 (C+ ) and the “Weak Paley–Wiener Theorem” This section is a somewhat technical portion of this chapter: here we transfer the space H 2 (D) into H 2 (C+ ) and H 2 (C+ ), and obtain descriptions of the subspaces invariant under dilations and under multiplication by characters.

6.3.1 A Unitary Mapping of L2 (T) onto L2 (R) Let C+ = {z ∈ C : Im(z) > 0} and ω : D → C+ the conformal mapping ω(z) = i

1+z . 1−z

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

200

RH, Dilations, and H 2 in the Hilbert Multi-disk

The restriction of the inverse w−i w+i on the boundary ∂C+ = R is a bijection R → T \ {1} whose Jacobian is ω−1 (w) =

|J(x)| =

2 , 1 + x2

x ∈ R.

Hence the mapping U,

 x − i 1 ·f U f (x) = √ , x+i π(x + i)

x ∈ R,

is a unitary isomorphism between the spaces L2 (T) and L2 (R), U : L2 (T) → L2 (R). Recall that in this chapter, a space L2 (E), where E ⊂ R, is regarded as a subspace of L2 (R) containing the functions of L2 (R) that are zero on the complement R \ E. (a) Lemma. UH 2 (D) = spanL2 (R)

  1 : Im(μ) > 0 . x−μ

Proof First observe that H 2 (D) = spanL2 (T)



1 1 − λz

 :λ∈D ;

the inclusion ⊃ is evident, and the converse follows from the fact that  1  f ∈ H 2 (D) and 0 = f, = f (λ) (∀λ ∈ D) 1 − λz imply f = 0. Then, clearly, U

1 1 − λz

=

cλ x−μ

where μ = ω(λ), and μ runs over the half-plane C+ as λ runs over D. Since  U : L2 (T) → L2 (R) is unitary, we obtain the stated equality. (b) Definition (the Hardy space H 2 (C+ ), inner functions). By definition, H 2 (C+ ) = UH 2 (D). A function θ in C+ is said to be inner (in C+ ) if θ ◦ω = Θ is an inner function in the disk D. Hence, θ is inner if and only if it is holomorphic and bounded, and

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.3 H 2 (C+ ) and the “Weak Paley–Wiener Theorem”

201

if its boundary values θ(x) = limy→0 θ(x + iy) (which exist by Fatou’s theorem and the conformal character of ω) are unimodular a.e. on R. (c) Lemma (semigroup of characters in L2 (R)). Let  z + 1 u s = exp s , s ∈ R, z−1 and let F ⊂ L2 (T) be a closed subspace. The following assertions are equivalent. (1) zF = F (respectively, zF ⊂ F). (2) u s F ⊂ F(∀s ∈ R) (respectively, ∀s ≥ 0). (3) eisx UF ⊂ UF (∀s ∈ R) (respectively, ∀s ≥ 0). Proof The equivalence (2) ⇔ (3) is clear, as is the implication (1) ⇒ (2) (seen in Exercise 1.8.3(a), and as u s ∈ H ∞ for s ≥ 0). To show (2) ⇒ (1), first observe that (2) implies ϕ s F ⊂ F for every s > 0, where us − 1 + s . ϕs = us − 1 − s Furthermore, for every ζ ∈ T, we have Re(1 − u s (ζ)) ≥ 0 and hence |ϕ s (ζ)| ≤ 1. Moreover, since e sw − 1 = sw + o(s) (∀w ∈ C) as s → 0, we obtain, for every ζ ∈ T\{1}, ϕ s (ζ) = ζ +o(1) (as s → 0). By the dominated convergence theorem,  for every f ∈ F, lim s→0 ϕ s f − z f 2 = 0, hence z f ∈ F. (d) Corollary. Let E ⊂ L2 (R) be a closed subspace. (1) eisx E ⊂ E, ∀s ∈ R ⇔ E = χA L2 (R) where A ⊂ R is a Borel set. (2) eisx E ⊂ E, ∀s ≥ 0 (and ∃s > 0 such that eisx E  E) ⇔ E = qH 2 (C+ ) where q is a unimodular function. (3) eisx E ⊂ E ⊂ H 2 (C+ ), ∀s ≥ 0, E  {0} ⇔ E = θH 2 (C+ ) where θ is an inner function in C+ . Indeed, to deduce (d) from (c) and the descriptions in Corollaries 1.4.1 and 1.4.3, it suffices to remark that if f ∈ L2 (T) and ϕ ∈ L∞ (T), then U(ϕ f ) = (ϕ ◦ ω)U f .  (e) The arithmetic of inner functions and canonical factorization in H2 (C+ ). Clearly the simple change of variables Θ −→ Θ ◦ ω−1 provides a bijective correspondence between the inner functions in D and those in C+ ; it allows the transfer of all the arithmetic of the inner functions of D (see § 3.2) to C+ (divisibility, GCD, LCM, spectrum, rules of calculus of the spectrum,

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

202

RH, Dilations, and H 2 in the Hilbert Multi-disk

etc.). Without unnecessary repetition of all these rules, we use them (with the modifications due to the change of variables) as needed. The question of canonical factorization is treated in the same manner: for f ∈ H 2 (C+ ) we write f = Ug = (gin ◦ ω−1 )Ugout , where g ∈ H 2 (D), and define fin = gin ◦ ω−1 , fout = Ugout . In particular, a singular function with singularity at the point z = 1,  1 + z Θ = exp −a , 1−z is transformed into θ(w) = Θ ◦ ω−1 = eiaw , w ∈ C+ (a function singular “at infinity”).

6.3.2 Fourier Transforms and the “Weak Paley–Wiener Theorem” Another unitary mapping, this time of L2 (R) onto L2 (R), is given by the Fourier transform and its inverse (Plancherel’s theorem: see Appendix A),   1 1 −ixz −1 f (x)e dx, F f (z) = √ f (x)eixz dx. F f (z) = √ 2π R 2π R The next lemma is a “weakened” form (but sufficient for our needs) of an important theorem of Paley and Wiener.

Raymond E. A. C. Paley (1907–1933) was a brilliant English mathematician, who studied at Eton and then at Trinity College, Cambridge, and was elected a Fellow in this prestigious establishment at the age of 23. As he was “inspired by the genius of G. H. Hardy and J. E. Littlewood” (in the words of Norbert Wiener), Paley worked mainly in harmonic analysis but also in probability and graph theory. Over the very short period of his professional career (not even three full years), he collaborated with a group of remarkable mathematicians, including Littlewood, Zygmund, Wiener, and P´olya. With Littlewood

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.3 H 2 (C+ ) and the “Weak Paley–Wiener Theorem”

203

he constructed the Littlewood–Paley decomposition – a tool of “hard” harmonic analysis that has become classical and indispensable in any serious application of Fourier analysis: to weighted singular integrals, Fourier multipliers, the maximal regularity of semigroups, etc. Another fundamental result is the Paley–Wiener theorem (1932, published posthumously) linking the Hardy space in the half-plane with the Fourier transform of L2 (R+ ) (see § 6.3.2 for a “weak form” of the result). According to Wiener, Paley was one of the pioneers in using probabilistic methods in harmonic analysis. He also discovered an interpolation of lacunary Fourier coefficients of a totally new kind, and contributed to graph theory. Paley’s name is linked with several important subjects, such as Littlewood–Paley theory, Paley–Wiener spaces, the Paley–Zygmund inequality, and Paley graphs. Paley was killed by an avalanche in the Canadian Rockies (near Calgary) at an altitude of 3 000 meters, during a skiing weekend. As Wiener wrote in his obituary for the Bulletin of the American Mathematical Society (vol. 39 (1933), p. 476), “the impression which Paley had made on American mathematicians is remarkable in the extreme . . . his premature death is an irreparable loss to mathematics.”

Pr {Z ≥ θE(Z)} ≥

(1 − θ)2 (E(Z))2 (1 − θ)2 (E(Z))2 + Var Z

Paley’s impact on mathematics is very diverse and goes far beyond analysis. Left, the order-13 Paley graph allowing graph-theoretic tools to be applied to the number theory of quadratic residues. Right, a useful Paley–Zygmund inequality which bounds the probability that a positive random variable is small, in terms of its mean and variance. (a) Lemma (Paley and Wiener, 1934). H 2 (C+ ) = F −1 L2 (R+ ) = F L2 (R− ), R+ = (0, ∞), R− = (−∞, 0).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

204

RH, Dilations, and H 2 in the Hilbert Multi-disk

Proof We calculate the inverse Fourier transform of an exponential function χR+ eiλx , λ ∈ C+ : 1 F −1 (χR+ eiλx )(z) = √ 2π



i 1 . eiλx eixz dx = √ · z + λ R+ 2π

Since F −1 is an isometric mapping, by the Lemma in § 6.3.1(a) it remains only to verify that the family χR+ eiλx , λ ∈ C+ generates L2 (R+ ), L2 (R+ ) = spanL2 (R) (χR+ eiλx : λ ∈ C+ ). To verify this last equality, suppose there is a function f ∈ L2 (R+ ) such that for every λ ∈ C+ , f ⊥ χR+ eiλx . Calculating with λ = i + y, y ∈ R, we obtain 0 = ( f, χR+ eiλx ) = F ( f χR+ e−x )(y) (∀y ∈ R), hence f χR+ e−x = 0, and thus f = 0.



(b) Lemma. For every s ∈ R, we have F eisx F −1 = τ s on the space L2 (R) where τ s is the operation of translation, τ s f (x) = f (x − s). Proof Direct computation.



(c) Corollary (Lax, 1959). Let E ⊂ L2 (R) be a (closed) subspace. (1) τ s E ⊂ E, ∀s ∈ R ⇔ E = F (χA L2 (R)) where A ⊂ R is a Borel set. (2) τ s E ⊂ E, ∀s ≥ 0 (and ∃s > 0 such that τ s E  E) ⇔ E = F (qH 2 (C+ )) where q is a unimodular function. (3) τ s E ⊂ E ⊂ L2 (R+ ), ∀s ≥ 0, E  {0} ⇔ E = F (θH 2 (C+ )) where θ is an inner function in C+ . The corollary is immediate by (a), (b) and § 6.3.1(d).



6.3.3 The Mellin Transform and the Group of Dilations The Mellin transform F∗ is the Fourier transform on the multiplicative group R+ = (0, ∞), that can be obtained with the aid of the group homomorphism ϕ : R → R+ , ϕ(x) = e−x :  dy 1 f (y)yz . F∗ f = F ( f ◦ ϕ), F∗ f (z) = √ y 2π R+

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.3 H 2 (C+ ) and the “Weak Paley–Wiener Theorem”

205

Robert Hjalmar Mellin (1854–1933) was a Finnish mathematician, a native of Liminka, a remote village of 3 000 inhabitants in the north of Finland. Son of a clergyman, Mellin studied at the University of Helsinki under Mittag-Leffler (submitting a thesis in 1881 on algebraic functions), and then continued in Berlin under the supervision of Weierstrass. After a few years at the University of Stockholm, he obtained a position as professor at the newly founded Technical University of Finland (1908). During this long episode of his career, Mellin made a noble gesture: in 1901, when applying for the post of professor in Helsinki, he withdrew to leave the place to the young prodigy Ernst Lindel¨of. Mellin is principally known for the study of an integral transform that bears his name (see § 6.3.3) and for his philosophical opposition to the theory of relativity, with at least ten articles published on this subject. In a long series of articles, he applied his transform to the study of the gamma and hypergeometric functions, Dirichlet series, the Euler ζ function and other arithmetic functions, and asymptotic developments. Mellin was one of the founders of the Finnish Academy of Sciences (1908) and was the Finnish representative on the editorial board of Acta Mathematica. Mellin was also known as an activist in the Fennoman movement, promoting the use of the Finnish language in place of Swedish, which had been dominant, especially before the attachment of Finland to the Russian Empire in 1809. Plancherel’s theorem states that F∗ is a unitary transform of L2 (R+ , dy/y) onto L2 (iR), with the same sense given to the integral as in the case of the group R (see Appendix A). As by tradition the definition of F∗ f is given on the imaginary axis iR, we “turn by π/2” the definition of the Hardy space, and examine H 2 (C+ ) = { f : f (z) = g(iz),

g ∈ H 2 (C+ )}.

With the notation above, as well as the evident equality L2 (R+ ) = L2 ((0, 1), dy/y) ◦ ϕ, we obtain the following version of the Paley–Wiener lemma. (a) Lemma (Paley and Wiener, 1934). H 2 (C+ ) = F∗ L2 ((0, 1), dy/y). (b) The group of dilations Dt f (y) = f (ty) (t > 0, y > 0)

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

206

RH, Dilations, and H 2 in the Hilbert Multi-disk

is a unitary group of transforms of the space L2 (R+ , dy/y) (verify!). We have    dy  dy  Dt L2 (0, 1), ⊂ L2 (0, 1), , t ≥ 1, y y   t−z dy 1 z dy = √ f (ty)y f (y)yz F∗ Dt f (z) = √ y y 2π R+ 2π R+ −z −sz = t F∗ f (z) = e F∗ f (z), where f ∈ L2 (R+ , dy/y), s = log(t) ∈ R. Hence, F∗ transforms a dilatation Dt into a multiplication by the character e−sz , s = log(t). Given the unitary property of F∗ , we obtain the following description of the subspaces invariant under dilations. (c) Corollary (subspaces invariant under dilations). Let E ⊂ L2 (R+ , dy/y) be a closed subspace. (1) Dt E ⊂ E, ∀t ∈ R+ ⇔ F∗ E = χA L2 (R), where A ⊂ R is a Borel set. (2) Dt E ⊂ E, ∀t ≥ 1 (and ∃t > 0 such that Dt E  E) ⇔ F∗ E = qH 2 (C+ ), where q is a unimodular function. (3) Dt E ⊂ E ⊂ L2 ((0, 1), dy/y), ∀t ≥ 1, E  {0} ⇔ F∗ E = θH 2 (C+ ), where θ is an inner function in C+ . The corollary is immediate by § 6.3.1(d) and/or § 6.3.2(c).



6.3.4 Completeness of the Characters, the Translations, and/or the Dilations By a simple comparison of the proposition in § 3.2.2(c) and of points § 6.3.1(d), § 6.3.2(c) and § 6.3.4(c) above, we obtain the following descriptions of the subspaces generated by a given family of functions Φ, Φ  {0}. (a) Let Φ ⊂ H 2 (C+ ) be a subset of H 2 (C+ ) and let   EΦ = spanH 2 (C+ ) eisx Φ : s ≥ 0 be the invariant subspace under (eisx ) s≥0 generated by Φ. Then, EΦ = θH 2 (C+ ) where θ = GCD(ϕin : ϕ ∈ Φ). (b) Let Φ ⊂ L2 (R+ ) and let   EΦ = spanL2 (R+ ) τ s Φ : s ≥ 0 be the subspace invariant under translations (τ s ) s≥0 generated by Φ. Then, EΦ = F (θH 2 (C+ )) where θ = GCD((F −1 ϕ)in : ϕ ∈ Φ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.4 The Nyman Theorem

207

(c) Let Φ ⊂ L2 ((0, 1), dy/y) and let

  EΦ = spanL2 (R+ ,dy/y) Dt Φ : t > 1

be the subspace invariant under dilations (Dt )t≥1 generated by Φ. Then F∗ EΦ = θH 2 (C+ ) where θ = GCD((F∗ ϕ)in : ϕ ∈ Φ).

6.4 The Nyman Theorem We are now ready to return to Theorem 6.2.1 and prove that the approximation by dilations mentioned in this theorem is not only sufficient, but is in fact equivalent to the Riemann hypothesis. Recall (see § 6.2): V = Lin(Dt ϕ : t >  1), V0 = f ∈ V : f (1) = 0 , where ϕ(x) = ρ(1/x) = 1/x − [1/x] (x > 0). Theorem 6.4.1 (Nyman, 1950) Let d = distL2 (0,1) (1, V0 ). The following assertions are equivalent. (1) (2) (3) (4)

d = 0. χ(0,1) ∈ closL2 (0,∞) (V). closL2 (0,∞) (V0 ) = L2 (0, 1). (RH) is correct (i.e. all the zeros of ζ in Re(s) > 0 are on the line Re(s) = 1/2).

Proof The implications (1) ⇔ (2) ⇔ (3) ⇒ (4) are already known (see Theorem 6.2.1 and Lemma 6.2.3). Let us show that (4) ⇒ (3). Denote F = closL2 (0,∞) (V0 ), and let V be the isometry L2 (0, 1) → L2 ((0, 1), dy/y) defined by V f (y) = y1/2 f (y). Since Dt V0 ⊂ V0 for t ≥ 1, we have Dt F ⊂ F and Dt (V F) ⊂ V F for every t ≥ 1. Also note that VV0 = Lin(Ψ), where  #  " 1 1/2 ϕ(ty) − ϕ(y) : t ≥ 1 ; Ψ = ψt := y t  ak 1/2  indeed, if f = y k ak ϕ(tk y) (tk > 1) and k tk = 0, then



ak f = y1/2 ak ϕ(tk y) = y1/2 (ak ϕ(tk y) − ϕ(y)) = ak ψtk . tk k k k By § 6.3.4(c), F∗ (V F) = θH 2 (C+ ),

where θ = GCD((F∗ ψt )in : it ≥ 1).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

208

RH, Dilations, and H 2 in the Hilbert Multi-disk

We calculate F∗ ψt using Lemma 6.1.3:  1  1 1 dy dy 1 y1/2 ϕ(ty)yz − √ y1/2 ϕ(y)yz F∗ ψt (z) = √ y y 2π 0 t 2π 0 " # 1 ζ(z + 1/2) ζ(z + 1/2) 1 1 − + − = √ 2π t(z − 1/2) (z + 1/2)tz+1/2 t(z − 1/2) t(z + 1/2)   1 ζ(z + 1/2) 1 1 − z+1/2 . = √ · z + 1/2 t t 2π To find (F∗ ψt )in , we use § 3.2 (transferred to the half-plane C+ with the aid of the conformal mapping ω = (1 + z)/(1 − z)). The roots of F∗ ψt (z) = 0 consist of the union   

1 1 = 0 ∪ z ∈ C+ : z−1/2 = 1 z ∈ C+ : ζ z + 2 t

   

2πin 1 1 = z ∈ C+ : ζ z + , n ∈ Z \ {0} . = 0 ∪ z ∈ C+ : z = + 2 2 log(t) Clearly the common zeros of the F∗ ψt , t ≥ 1 (and hence of θ), are exactly

  1 Z = z ∈ C+ : ζ z + =0 . 2 Each function F∗ ψt is holomorphic at the boundary z ∈ iR, thus the singular measure of (F∗ ψt )in ◦ ω is situated at the point {1}. However the logarithmic residue vanishes: log |F∗ ψt (x)| log |1/(x + 1/2)| = lim = 0, lim x>0,x→∞ x>0,x→∞ x x hence the singular part of (F∗ ψt )in is trivial (constant). Conclusion The function θ is a Blaschke product corresponding to the set of the zeros Z (more exactly, θ ◦ ω is a Blaschke product in the disk D corresponding to the zeros ω−1 (Z)), hence F∗ (V F) = H 2 (C+ ) if and only if Z = ∅. 

6.5 The Distance Function and Zero-free Disks of ζ We introduce here a family of approximation problems which generalize Nyman’s approach in § 6.4 and provide a number of disks in C+ free of zeros of the Euler ζ function. We begin with a few observations on the reproducing kernel of the space H 2 (D), and then transfer them to H 2 (C+ ) and L2 ((0, 1), dy/y) to obtain the result.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.5 The Distance Function and Zero-free Disks of ζ

209

6.5.1 The Distance Function Let λ ∈ D; the reproducing kernel at the point λ is the function kλ , kλ (z) =

k≥0

1

k

λ zk =

1 − λz

(z ∈ D),

such that ϕλ ( f ) = ( f, kλ )H 2 for every function f ∈ H 2 , where ϕλ is an evaluation functional at the point λ, ϕλ ( f ) = f (λ),

f ∈ H2

(see also Exercise 2.8.2(b) where kλ is mentioned). Let E ⊂ H 2 be a subspace and ΘE (λ) = ϕλ |E,

dE (λ) = distH 2 (kλ , E),

λ ∈ D.

The following properties are evident: (a) (b) (c) (d)

kλ (z) = 1/(1 − λz) and ΘH 2 (λ) = kλ  = (1 − |λ|2 )−1/2 . ΘE (λ) = dE ⊥ (λ), where E ⊥ is the orthogonal complement of E in H 2 . Θ2E (λ) + dE2 (λ) = kλ 2 = (1 − |λ|2 )−1 . If E = ΘH 2 , where Θ is an inner function, then ΘE (λ) = |Θ(λ)|(1 − |λ|2 )−1/2 .

Lemma 6.5.1 Let E ⊂ H 2 (D) be a subspace, Z(E) = {z ∈ D : f (z) = 0 ∀ f ∈ E} (the common zeros of E), λ ∈ D and  Dλ = z ∈ D : |bλ (z)|2 < 1 − dE2 (λ)/kλ 2 , where bλ is the elementary Blaschke factor. Then, Z(E)



Dλ = ∅.

Proof Let z ∈ Z(E) and f ∈ E. Then, f = bz g where g ∈ H 2 ,  f  = g. Consequently, | f (λ)| = |bz (λ)| · |g(λ)| ≤ |bz (λ)| · kλ  ·  f , and then ΘE (λ) ≤ |bz (λ)| · kλ . Thus z  Dλ , since 1 − dE2 (λ)/kλ 2 = Θ2E (λ)/kλ 2 . Theorem 6.5.2 Let s ∈ C+ , γ > 0,     1 Ψγ = Lin ψt,γ (y) := yγ ϕ(ty) − ϕ(y) : t ≥ 1 , t and dγ (s) = distL2 (0,1;dy/y) (x s , Ψγ ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press



210

RH, Dilations, and H 2 in the Hilbert Multi-disk

Then the disk D s,γ ,

 z − s 2   < 1 − 2 Re(s)d2 (s) , D s,γ = γ + D s = γ + z :  γ  z − s∗ is free of zeros of the Euler ζ function; here s∗ designates the point of C symmetrical with s with respect to iR.

g

t

g

s g

s s = s + it

s

g

g +s

A shifted non-Euclidean disk (shaded) from Theorem 6.5.2 free of zeros of the ζ function (γ and s, the non-Euclidean center, are free parameters).

Proof We begin by noting that the case where γ = s = 1/2 corresponds to Theorem 6.4.1. We re-use the computation of Theorem 6.4.1 by replacing 1/2 with γ > 0. Namely, in the same manner as in Theorem 6.4.1, we have 1  1 ζ(z + γ)  1 − z+γ . F∗ ψt,γ (z) = √ · z+γ t t 2π Consequently, setting Fγ = closL2 (0,1;dy/y) (Ψγ ), we obtain

 Z(F∗ Fγ ) = z ∈ C+ : ζ(z + γ) = 0 .

Moreover, the Mellin transform F∗ : L2 ((0, 1), dy/y) → H 2 (C+ ) is isometric, and dγ (s) = distH 2 (C+ ) (F∗ x s , F∗ Fγ ) where F∗ x s (z) = π−1/2 (s + z)−1 (z ∈ C+ ) is the reproducing kernel of H 2 (C+ ), F∗ x s 2H 2 (C+ ) = x s 2L2 (0,1,dy/y) = 1/2 Re(s). By Lemma 6.5.1 (transferred to C+ ), the disk

 z − s 2   < 1 − 2 Re(s)d2 (s) z :  γ z − s∗  is free of zeros of F∗ Fγ , and the result follows.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press



6.6 Completeness of Dilations and the Hilbert Multi-disk

211

Theorem 6.5.3 Let γ > 0. With the notation of Theorem 6.5.2, the following assertions are equivalent. (1) The function ζ does not vanish in the half-plane {z : Re(z) > γ}. (2) Fγ = L2 ((0, 1), dy/y). (3) x s ∈ Fγ (∃s ∈ C+ , or ∀s ∈ C+ ). Proof The implications (2) ⇒ (3) ⇒ (1) follow immediately from Theorem 6.5.2 (with d s (γ) = 0). For (1) ⇒ (2) we repeat the portion (4) ⇒ (3) of Theorem 6.4.1, but  replacing ψt by ψt,γ , ζ(z + 1/2) by ζ(z + γ), etc.

6.6 Completeness of Dilations and the Hilbert Multi-disk In this section, we study an approximation problem of the integer dilations Dn f (x) = f (nx), n = 1, 2, . . . , of an arbitrary function f , associated with the completeness (already treated) of all the dilations ϕ(tx), t ≥ 1. A link with the Riemann hypothesis is established in the following theorem, whose proof lies beyond the elementary framework of this chapter; this theorem shows that in Nyman’s Theorem 6.4.1 we can limit ourselves to the integer dilations Dn ϕ, n = 1, 2, . . . Theorem (B´aez-Duarte, 2003) The following assertions are equivalent. (1) (RH) is correct (i.e. all the zeros of ζ in Re(s) > 0 are on the line Re(s) = 1/2). (2) χ(0,1) ∈ closL2 (0,∞) (N), where N = Lin(ϕ(nx) : n = 1, 2, . . . ), ϕ(x) = ρ(1/x) = 1/x − [1/x] (x > 0).  (3) 1 ∈ closL2 (0,1) (N0 ), where N0 = f ∈ N : f (1) = 0 . Next, we examine the general question of cyclic vectors of the semigroup of dilations (Dn )n≥1 in the space L2 (0, 1).

6.6.1 The Wintner–Beurling Problem Aurel Wintner (1944) and, independently, Arne Beurling (1945) posed the 2 problem of the description of functions f ∈ L2 (0, 1) such that E D f = L (0, 1), where ED f = spanL2 (0,1) (Dn f : n = 1, 2, . . . ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

212

RH, Dilations, and H 2 in the Hilbert Multi-disk

A function with such a completeness property is called (Dn )-cyclic. Of course, it is necessary to describe precisely how f is defined outside the interval (0, 1). We have selected (following Wintner and Beurling) the extension given by the development of f over the orthonormal trigonometric basis

√ f (x) = ak 2 sin(πkx), x ∈ (0, 1). k≥1

The choice of this basis is justified by the fact that this is the only orthonormal basis of L2 (0, 1) of the form (Dn e)n≥1 , where e ∈ L2 (0, 1) (see Exercise 6.7.2). Hence, a function f ∈ L2 (0, 1) extends to R in an odd and 2-periodic manner. Given the result cited above, the Wintner–Beurling problem is clearly linked to the Riemann hypothesis. It remains unsolved to this day (2018).

Aurel Wintner (second on the left in the third and last row) during a seminar at the Niels Bohr Institute in Copenhagen in 1930. In the first row (where four Nobel Prize winners appear) are Oscar Klein, Niels Bohr, Werner Heisenberg, Wolfgang Pauli, George Gamow, Lev Landau, and Hendrik Kramers. Edward Teller is the first on the right in the second row. Aurel Wintner (1903–1958), a Hungarian–American mathematician, was one of the principal promoters of analysis of the twentieth century, whose heritage is without doubt not yet properly recognized. Wintner published 437 articles (!) in the most renowned journals, as well as nine reference texts, including Spektraltheorie der unendlichen Matrizen: Einf¨uhrung in den analytischen Apparat der Quantenmechanik (1929),

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.6 Completeness of Dilations and the Hilbert Multi-disk

213

Analytical Foundations of Celestial Mechanics (1941, 1947), The Fourier Transforms of Probability Distributions (1947), and An Arithmetical Approach to Ordinary Fourier Series (1945). It was in fact the first of these monographs that provided the very first rigorous treatment of spectral theory: in particular, the spectral theorem for the normal operators, the fact that the spectrum of a bounded operator is non-empty (long before Gelfand theory), the essential spectrum, etc. The second in the list is a standard reference for the subject. Wintner was one of the founders (with Paul Erd˝os and Mark Kac) of probabilistic number theory (1940): he introduced the expressions law of the iterated logarithm, essential spectrum, summing method of Eratosthenes, etc. (The latter was reinvented independently – but later – by Albert Ingham and introduced into common usage by G. H. Hardy under the erroneous name “Ingham’s method.”) Wintner was born in Budapest, and after a long hesitation between the sciences and music (he showed a rare talent for the violin), he registered at the University of Budapest (1920) for studies in astronomy and physics. He was forced to leave in 1924 because of the galloping inflation that induced chaos in the finances of his family (and the country). During the following three years (hence, between the ages of 21 and 23) he published about 20 articles in prestigious journals (in celestial mechanics and mathematics). He obtained his doctorate under the supervision of Leon Lichtenstein (Leipzig, 1927), then collaborated with Levi-Civita in Rome and Str¨omberg in Copenhagen. Next, Wintner moved to the USA (Princeton, Harvard, MIT, and then the Johns Hopkins University) to work with Birkhoff, Erd˝os, Kac, and Wiener. In particular, he was the very first to propose the use of the approximation technique of harmonic analysis for problems in arithmetic (partially presented in § 6.6). Editor of the American Journal of Mathematics, Wintner played a principal role (with Andr´e Weil) in giving this journal its dominant stature and creating its irreproachable standards. During a meeting of the American Mathematical Society (Columbus, 1940), Wiener and Wintner invented a satirical “journal,” Trivia Mathematica, and amused themselves by proposing titles of articles deemed “acceptable.” In 1930 Wintner married Irmgard H¨older (daughter of Otto H¨older) and had one child. He died suddenly of a heart attack while he was at the peak of his mathematical productivity.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

214

6.6.2 Change of Orthonormal Basis: The Semigroup T = (T n) on H 2 0

Clearly, nothing would be changed if we replace the series

√ f = ak 2 sin(πkx) k≥1

with  f 22 =



|ak |2

k≥1

and the semigroup Dn f =



√ ak 2 sin(πknx)

k≥1

with the power series f =



ak zk

k≥1

with  f 22 =

 k≥1

|ak |2 < ∞ and the semigroup T = (T n )n≥1 , T n f (z) = f (zn ).

Then the Wintner–Beurling problem becomes a problem of the functions f ∈ H02 = { f ∈ H 2 (T) : f (0) = 0}, cyclic with respect to the semigroup T , i.e. such that E Tf = H02 , where E Tf = spanH 2 (T) (T n f : n = 1, 2, . . . ). Clearly (as with the dilations (Dn )) (T n ) is a multiplicative semigroup (or more precisely, a representation n −→ T n of N in H02 ): T n T m = T mn .

6.6.3 The Reproduction of Variables and the Bohr Transform The decomposition into prime numbers n = pα1 1 pα2 2 . . . pαk k , where α j ∈ Z+ =  N {0} and k = k(n) (see § 6.1.1), gives T n = T pα11 · · · T pαkk , and suggests considering z p1 , . . . , z pk , . . . as independent variables.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.6 Completeness of Dilations and the Hilbert Multi-disk

215

To formalize this idea we will need some new notation. Namely, for a natural number n = pα1 1 pα2 2 · · · pαs s represented in its Euclidean decomposition, we associate a unique multi-index α(n) = (α1 , α2 , α3 , . . . ), 

where α j ∈ Z+ = N {0} (and after a certain rank α j = 0). In fact, the mapping α : n −→ α(n) is a bijection of N onto the set  Zk+ Z+ (∞) = k≥1

of the finitely supported sequences of non-negative integers, and moreover, this is a homomorphism of the multiplicative semigroup N in Z+ (∞). To shorten the   notation with the multi-indices we systematically replace α∈Z+ (∞) with α≥0 . Now, we can define what is known as the Bohr transform: for

fˆ(n)zn ∈ H02 , f = n≥1

set U f (ζ) =



fˆ(n)ζ α(n) ,

n≥1

where ζ = (ζ1 , ζ2 , . . . ) ∈ D∞ is such that the series converges absolutely, and a multi-power ζ α , defined by ζ α = ζ1α1 ζ2α2 · · · ζ sαs · · · . This is a finite product because the multi-index α has finite support. For every series U f (ζ) to be well-defined at the point ζ = (ζ1 , ζ2 , . . . ) ∈ D∞ , it is necessary and sufficient that

|ζ α |2 < ∞ α≥0



(indeed, to have α≥0 |aα bα | < ∞ for every a = (aα ) ∈ l2 it is necessary and  sufficient that α≥0 |bα |2 < ∞: see Appendix A). Lemma 6.6.1 Let ζ = (ζ1 , ζ2 , . . . ) ∈ D∞ . Then

 1 |ζ α |2 = , 1 − |ζk |2 α≥0 k≥1 and hence



α≥0

|ζ α |2 < ∞ if and only if

 k≥1

|ζk |2 < ∞.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

216

Proof We repeat (almost) the proof of § 6.1.2: for a series of positive terms we can write



|ζ α |2 = lim |ζ α |2 , k

α≥0



|ζ α |2 =





α2 ≥0...αk ≥0 α1 ≥0

α∈Zk+

=

α2 ≥0...αk ≥0

α∈Zk+

|ζ1α1 |2 |ζ2α2 . . . ζkαk |2

(1 − |ζ1 |2 )−1 |ζ2α2 . . . ζkαk |2 =

k  (1 − |ζ j |2 )−1 , j=1



and the result follows.

6.6.4 The Hilbert Multi-disk D∞ and the Space H 2 (D∞ ) 2

2

In 1909, Hilbert published a sketch of the theory of infinitely many complex variables in a Hilbert multi-disk D∞ 2 , 2 D∞ 2 = {ζ = (ζ1 , ζ2 , . . . ) : ζ ∈ l , |ζ j | < 1 for all j ≥ 1}.

We will need an analog of the Hardy space in D∞ 2 , which we define as the space of power series with coefficients in l2 :





α 2 2 H 2 (D∞ ) := F = c (F)ζ : F = |c (F)| < ∞ . α α 2 2 α∈Z+ (∞)

α∈Z+ (∞)

Lemma 6.6.2 (1) For every function f ∈ H02 and ζ ∈ D∞ 2 , the series U f (ζ) = converges absolutely, and   1 1/2 . |U f (ζ)| ≤  f 2 1 − |ζk |2 k≥1

 n≥1

fˆ(n)ζ α(n)

(2) U is a unitary mapping of H02 onto the Hardy space of the multi-disk, U : H02 → H 2 (D∞ 2 ), transforming the orthonormal basis (zn )n≥1 of H02 to the orthonormal basis of multi-powers (ζ α )α≥0 of H 2 (D∞ 2 ). 2 ∞ (3) For every function f ∈ H0 , ζ ∈ D2 and n ∈ N, (UT n f )(ζ) = ζ α(n) (U f )(ζ), where α(n) = (α1 (n), . . . , αk (n), . . . ) is defined in § 6.6.3.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.6 Completeness of Dilations and the Hilbert Multi-disk

217

(4) Lat(T n ) = U −1 Lat(Mζ ), which means that a closed subspace E ⊂ H02 is T -invariant (i.e. T n -invariant for every n ∈ N) if and only if UE is Mζ -invariant (i.e. ζk -invariant for every k ∈ N: f ∈ UE ⇒ ζk f ∈ UE for every k ∈ N). (5) A function f ∈ H02 is T -cyclic if and only if U f is Mζ -cyclic in H 2 (D∞ 2 ), i.e. EU f = H 2 (D∞ 2 ),

where EU f := span(ζ α U f : α ∈ Z+ (∞))

and span denotes the closed linear hull in H 2 (D∞ 2 ). Proof (1) By Cauchy–Schwarz and Lemma 6.6.1,

 |ζ α |2 =  f 22 |U f (ζ)|2 ≤  f 22 α≥0

k≥1

1  . 1 − |ζk |2

(2) By definition, Uzn = ζ α(n) , for every n ∈ N. Moreover, α is a bijection of N on Z+ (∞), and the result follows. (3) Since α is a homomorphism, for every k ∈ N, we have (UT n zk )(ζ) = (Uzkn )(ζ) = ζ α(kn) = ζ α(n) ζ α(k) = ζ α(n) (Uzk )(ζ). Moreover, the functional f −→ U f (ζ) is linear and bounded on H02 (see (1)), which implies the result. (4) Evident by (3). (5) Evident by (4).  Corollary 6.6.3 (Beurling, 1945) If f ∈ H02 is T -cyclic, then U f (ζ)  0 for every ζ ∈ D∞ 2 . Indeed, if U f (ζ) = 0, then for every α ∈ Z+ (∞), ζ α (U f )(ζ) = 0, and by  Lemma 6.6.2(1) g(ζ) = 0 for every g ∈ EU f . Hence, 1  EU f .

6.6.5 A Few Initial Observations In general, the necessary condition of Corollary 6.6.3 is not sufficient for the Mζ -cyclicity of U f in H 2 (D∞ 2 ) (as we have already seen in the case of the Mz -cyclicity in the space H 2 (D), for example, see Exercise 1.8.3(b)). In what follows, we will see that for certain classes of functions f the converse of Corollary 6.6.3 is nonetheless correct. But first, we make a few technical preparations.

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

218

RH, Dilations, and H 2 in the Hilbert Multi-disk

(a) The space H 2 (Dn ) in the polydisk Dn is defined as the space of power series in z = (z1 , . . . , zn ) ∈ Cn with coefficients in l2 :





H 2 (Dn ) := F = cα (F)ζ α : F22 = |cα (F)|2 < ∞ . α∈Zn+

α∈Zn+

Another description: H 2 (Dn ) is the subspace of H 2 (D∞ 2 ) consisting of the functions F depending only on the variables ζ1 , . . . , ζn (more precisely, consisting of the functions F ∈ H 2 (D∞ 2 ) such that cα (F) = 0 for any α = (α1 , . . . , α j , . . . ) such that α j > 0 for an index j > n). It is easy to see that for every function F ∈ H 2 (Dn ) (and with 0 < r < 1), 

2 2|α| 2 F2 = lim r |cα (F)| = lim |F(rζ)|2 dmn (ζ), r→1

r→1

α∈Zn+

Tn

where mn = m × m × · · · × m is the Lebesgue measure on the circle Tn . (b) An integral formula for the norm. Let F ∈ H 2 (D∞ 2 ). Then, by (a), 

F22 = lim |cα (F)|2 = lim lim |F(rζ)|2 dmn (ζ). n→∞

n→∞ r→1

α∈Zn+

Tn

(c) Reproducing kernel. By Lemma 6.6.2(1), for every λ ∈ D∞ 2 , the mapping F −→ F(λ) is a continuous linear functional on H 2 (D∞ 2 ), and hence there exists a unique ) such that function Kλ ∈ H 2 (D∞ 2 F(λ) = (F, Kλ ) reproducing kernel of H 2 (D∞ for every function F ∈ H 2 (D∞ 2 ). It is called the  2 ) ∞ at the point λ ∈ D2 . It is easy to find: F(λ) = α∈Z+ (∞) cα (F)λα , and thus, by uniqueness,

α λ ζ α , ζ ∈ D∞ Kλ (ζ) = 2 . α∈Z+ (∞)

The series converge absolutely, and hence by taking successive summations over α1 , α2 , etc. we obtain  1 (z ∈ C). Kλ (ζ) = kλ j (ζ j ) where kλ j (z) = 1 − λ jz j≥1  ∞ (d) Fact. H 2 (D∞ 2 ) = span Kλ : λ ∈ D2 , where the span is taken in the space H 2 (D∞ 2 ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.6 Completeness of Dilations and the Hilbert Multi-disk

219

∞ Proof If F ∈ H 2 (D∞ 2 ) and F ⊥ Kλ for every λ ∈ D2 , then F(λ) = 0 (∀λ), hence F = 0. 

(e) Weak convergence in H 2 (D∞ 2 ). A sequence (F n ) converges weakly to F ∈ ∞ ) if and only if sup F  H 2 (D∞ n 2 < ∞ and, for every λ ∈ D2 , limn F n (λ) = n 2 F(λ). 

Proof Clear by the Banach–Steinhaus theorem. (f) The space H ∞ (D∞ 2 ) is, by definition,

 2 ∞ H ∞ (D∞ 2 ) = F ∈ H (D2 ) : F∞ = sup |F(ζ)| < ∞ . ζ∈D∞ 2

2 ∞ It is clear by (b) above that, for all functions G ∈ H ∞ (D∞ 2 ), F ∈ H (D2 ), we 2 ∞ have FG ∈ H (D2 ) and

FG2 ≤ F2 G∞ . The property below will be useful in what follows. Fact For every function G ∈ H ∞ (D∞ 2 ), there exists a sequence of polynomials (pn ) such that pn ∞ ≤ G∞ and limn pn (λ) = G(λ), λ ∈ D∞ 2 . Proof Let G(n) (λ) =



cα (G)λα

α∈Zn+

 be the restriction of G to Dn . Since the series G(λ) = α∈Z+ (∞) cα (G)λα converges absolutely, we have limn G(n) (λ) = G(λ) for every λ ∈ D∞ 2 , and (of course) G(n) ∞ ≤ G∞ . We define the pn as the Fej´er polynomials of G(n) of degree N = N(n) sufficiently large: pn = G(n) ∗ ΦN,n , where ΦN,n is the Fej´er kernel on Tn (the product of the Fej´er kernels on T: see Appendix A). Clearly, for every n, the restrictions of the Fej´er approximation f −→ ( f − ΦN,n ∗ f ), N = 1, 2, . . . converge uniformly on the compact subsets n of Dn : precisely, for any compact subset ΔD , 0 < Δ < 1, the mappings ΦN,n,Δ : i f −→ ( f − ΦN,n ∗ f )|ΔDn tend to zero for the operator norm H 2 (Dn ) → H ∞ (ΔDn ), i.e. lim ΦN,n,Δ  = lim sup  f − ΦN,n ∗ f H ∞ (ΔDn ) = 0.

N→∞

N→∞  f 2 ≤1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

220

RH, Dilations, and H 2 in the Hilbert Multi-disk

Using this property, we obtain, for pn (ζ) = ΦN,n ∗ G(n) (ζ), |G(ζ) − pn (ζ)| ≤ |G(ζ) − G(n) (ζ)| + |G(n) (ζ) − ΦN,n ∗ G(n) (ζ)| ≤ G − G(n) 2 Kζ 2 + ΦN,n,Δ  · G(n) 2 , where Δ = Δ(ζ) = max j |ζ j | < 1. Now, clearly there exists a sequence N = N(n) → ∞ such that limn |G(ζ) − pn (ζ)| = 0 for every ζ ∈ D∞ 2 , uniformly on : K  ≤ A, Δ(ζ) ≤ Δ < 1}.  the sets {ζ ∈ D∞ ζ 2 (g) The invariant subspaces E F = span(ζ α F : α ∈ Z+ (∞)) generated in the 2 ∞ space H 2 (D∞ 2 ) by a function F ∈ H (D2 ) satisfy the following property: H ∞ (D∞ 2 ) · F ⊂ EF .

Proof Indeed, for a function G ∈ H ∞ (D∞ 2 ) and with the polynomials pn of (f), we have pn F2 ≤ pn ∞ F2 ≤ G∞ F2 and limn pn (λ)F(λ) = G(λ)F(λ), λ ∈ D∞ 2 . By (e) above, we have limn pn F = GF for the weak convergence of 2 ∞  H (D2 ), and hence FG ∈ E F .

6.6.6 Cyclic Polynomials We now show that the necessary condition of Corollary 6.6.3 is also sufficient for the cyclicity of polynomials. Remark that a function f ∈ H02 is a polynomial (in D) if and only if its image U f (after a reproduction of the variables) is a 2 ∞ polynomial in D∞ 2 . See § 6.6.5 for a larger class of functions F ∈ H (D2 ) for which the same description remains correct. (a) Theorem (Neuwirth, Ginsberg, and Newman, 1970). Let f be a polynomial in H02 . The following assertions are equivalent. (1) f is T -cyclic in H02 . (2) U f is Mζ -cyclic in H 2 (D∞ 2 ). (3) U f (ζ)  0 for every ζ ∈ D∞ 2 . For the proof, we will need the following lemma, itself of independent interest. (b) Lemma (Neuwirth, Ginsberg, and Newman, 1970). Let F be a polynomial ∞ in D∞ 2 such that F(ζ)  0 for every ζ ∈ D2 . Then    F(ζ)  ≤ 2deg(F)  F(rζ) 

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.6 Completeness of Dilations and the Hilbert Multi-disk

221

n for every 0 ≤ r ≤ 1 and ζ ∈ D∞ 2 . The example F(z) = (1 + z) shows (for z → 1 and r → 0) that this upper estimate is sharp.

Proof Since a polynomial depends on only a finite number of the variables ζ1 , ζ2 , . . . (say m variables), we can restrict ourselves to ζ ∈ Dm . First suppose that m = 1. Then F(z) = A(z − z1 ) . . . (z − zd ) where |z j | ≥ 1. If z ∈ D, then |z − rz| ≤ 1 − r ≤ |z j | − r|z| ≤ |rz − z j |, and hence      z − z j  = 1 + z − rz  ≤ 2.  rz − z j   rz − z j  In the general case, we take ζ = (ζ1 , . . . , ζm ) ∈ Dm and consider the polynomial in a single variable Pζ (z) = F(zζ), z ∈ D. Applying the inequality already proved for m = 1, we obtain    Pζ (z)  ≤ 2deg(F) for all z ∈ D.  Pζ (rz)  It only remains to select z = 1.



(c) Proof of Theorem (a) Clearly (1) ⇔ (2) ⇒ (3). To prove (3) ⇒ (2), denote F = U f . Since F depends on a finite number of the variables ζ j , j ≥ 1, we have, for any r, 0 < r < 1, inf |F(rζ)| > 0,

ζ∈D∞ 2

hence 1/Fr ∈ H ∞ (D∞ 2 ), and then F/F r ∈ E F , by § 6.6.5(g). By Lemma (b), F ≤ F ≤ 2deg(F) for all 0 < r < 1, Fr 2 Fr ∞ and thus limr→1 (F/Fr ) = 1 weakly in H 2 (D∞ 2 ) (see § 6.6.5(e)). Consequently, 2 ∞  1 ∈ E F , and hence E F = H (D2 ).

6.6.7 Other Classes of (T n)-cyclic Functions of H 2 0

We conclude this chapter with the statement of a theorem proved in Nikolski (2012) and examine certain consequences. We will deduce from it that every reproducing kernel Kλ is a cyclic function in H 2 (D∞ 2 ) and, in particular, the  functions fγ = k≥1 k−γ zk (Re(γ) > 1/2) are (T n )-cyclic in H02 , or, equivalently,  the functions ϕγ = k≥1 k−γ sin(πkx) (Re(γ) > 1/2) are (Dn )-cyclic in L2 (0, 1) (Wintner, 1944). Recall (see Lemma 6.6.2) that a function f ∈ H02 is (T n )cyclic in H02 if and only if U f is Mζ -cyclic in H 2 (D∞ 2 ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

222

RH, Dilations, and H 2 in the Hilbert Multi-disk

∞ (a) Theorem. Let F be a function of H 2 (D∞ 2 ) such that for every ζ ∈ D2 , F(ζ)  0. Each of the following conditions implies the Mζ -cyclicity of F in H 2 (D∞ 2 ).

(1) 1/F ∈ H ∞ (D∞ 2 ). (2) There exist > 0 and N ≥ 1 such that F 1+ ∈ H 2 (D∞ 2 ) and ). 1/F 1/N ∈ H 2 (D∞ 2 (3) Re(F(ζ)) ≥ 0 for every ζ ∈ D∞ 2 . (4) F depends on only a finite number of variables ζ j , 1 ≤ j ≤ m, and F ∈ Hol(rDm ), where r > 1. (5) F = U f where f ∈ H02 (D) has its Fourier spectrum σ( f ) = {k ∈ N : fˆ(k)  0} in a finitely generated multiplicative semigroup, and fˆ(k) = o(k− ) as k → ∞ ( > 0). (6) F = U f , f ∈ H02 with σ( f ) ⊂ {nk : k ∈ Z+ }, n > 1, and the function  ϕ = k≥0 fˆ(nk )zk is a Beurling outer function. (In this special case, the last condition is also necessary for the cyclicity.) (b) Corollary. Each reproducing kernel Kλ , λ ∈ D∞ 2 , is (Mζ )-cyclic in ), or, equivalently, every function H 2 (D∞ 2

fλ = λα(n) zn n≥1

is (T n )-cyclic in

H02 (D).

Indeed,



Kλ (ζ) =

α

λ ζα =

α∈Z+ (∞)



kλs (ζ s )

s≥1

  where ka (z) = 1 − az −1 (a, z ∈ D). Moreover, ka Hp p (T) = 1 + |pa/2|2 (1 + o(1)) when a → 0 (∀p < ∞; see Exercise 6.7.3), and hence  kλs Hp p (T) < ∞ for all λ, λ = (λ s ) ∈ D∞ Kλ  pp = 2 . s≥1

Similarly, 1/Kλ 22 =



1 − λ s ζ s 2H 2 (T) < ∞.

s≥1

Then the cyclicity of Kλ follows from Theorem (a), point (2). (c) Corollary (Wintner, 1944). Every function

k−a zk , Re(a) > 1/2 fa = k≥1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press



6.7 Exercises

223

is (T n )-cyclic in H02 . Indeed, U fa = Kλ where λ = (λ s ) s≥1 , λ s = p−a s (p s are primes).



6.7 Exercises 6.7.1 Multipliers of the Space H 2 (D∞ ) 2

The space of multipliers of H

2

(D∞ 2 )

is defined by  2 ∞ 2 ∞ Mult(H 2 (D∞ 2 )) := ϕ : F ∈ H (D2 ) ⇒ ϕF ∈ H (D2 )

equipped with the operator norm. ∞ ∞ Show that Mult(H 2 (D∞ 2 )) = H (D2 ) and that the norms coincide: ϕMult = ϕ∞ .

Solution: We have already mentioned at the beginning of § 6.6.5(f) that every function ϕ ∈ H ∞ (D∞ 2 ) is a multiplier with ϕMult ≤ ϕ∞ . The converse is true in general: a multiplier on a function space is bounded (whatever the set where the space is defined). In our case, we have |ϕ(λ)| = |(ϕn · 1)(λ)|1/n ≤ (ϕnMult 12 Kλ 2 )1/n ∞ ∞ for every λ ∈ D∞ 2 , which implies ϕ ∈ H (D2 ) and ϕ∞ ≤ ϕMult .



6.7.2 Orthogonal Dilations Here, we consider the functions f of the space L2 (0, 1) extended to R so as to be odd and 2-periodic (as in § 6.6). Prove the following properties. (1) Let g ∈ L2 (0, 1). The sequence (Dn g)n≥1 is an orthogonal basis of L2 (0, 1) if and only if g(x) = a · sin(πx), a  0. (2) Let f ∈ H02 . The sequence (T n f )n≥1 is an orthogonal basis of H02 if and only if f = a · z, a  0. α (3) Let F ∈ H 2 (D∞ 2 ). The family (ζ F)α∈Z+ (∞) is an orthogonal basis of 2 ∞ H (D2 ) if and only if f = constant = a, a  0. Solution: As explained in § 6.6, the three sequences are unitarily equivalent (with √ the correspondences 2 sin(πnx) → zn → ζ α(n) ), hence it suffices to show (3). As multiplication by ζ α is an isometry on H 2 (D∞ 2 ), we can suppose that F2 = 1. Then the orthonormal bases (ζ α ) and (ζ α F) are unitarily equivalent, which means

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

224

that the multiplication h −→ hF is a unitary mapping of H 2 (D∞ 2 ) onto itself. By ∞ ∞ Exercise 6.7.1, F ∈ H (D2 ) and F∞ = 1. However, the inverse of this mapping is also a unitary multiplier, thus 1/F∞ = 1. By the maximum principle applied to F(n) (see the definition in § 6.6.5(f)), we obtain consecutively F(1) = constant, F(2) = constant, etc., and clearly all these constants coincide. Since limn F − F(n) 2 = 0, we have F = constant. 

p

6.7.3 Asymptotics of k a  p as a → 0 Show that for every p > 0,   2 dm p  pa  (1 + (o(1)), = 1 + ka  p =  2  p T |1 − az|

as a → 0.

Solution: Recall that for any γ ∈ C and w → 0, we have (1 + w)γ = 1 + γw +

γ(γ − 1) 2 w (1 + o(1)). 2

Applying this with γ = −p/2 and |1 − az|2 = 1 + (|a|2 − 2 Re(az)), and taking into account   Re(az) dm = 0, (Re(az))2 dm = |a|2 /2, T

we obtain ka  pp =

T

 T

(1 + γ(|a|2 − 2 Re(az)) + (γ(γ − 1)/2)(|a|2 − 2 Re(az))2 (1 + o(1))) dm

=1+

 T

(γ|a|2 + (γ(γ − 1)/2))4|a|2 /2)(1 + o(1)) dm

 2 pa = 1 + γ2 |a|2 (1 + o(1)) = 1 +   (1 + (o(1)). 2



6.7.4 Particular Features of the Multi-disk D∞ 2

∞ ∞ 1 (a) Let λ = (λ1 , λ2 , . . . ) ∈ D∞ 2 . Show that Kλ ∈ H (D2 ) if and only if λ ∈ l .

Solution: By § 6.6.7(b), Kλ (ζ) =



α

λ ζα =

α∈Z+ (∞)

hence Kλ H ∞ (D∞2 ) ≤

 s≥1

kλs H ∞ (D) =





kλs (ζ s ),

s≥1

(1 − |λ s |)−1 ,

s≥1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.7 Exercises

Kλ H ∞ (D∞2 ) ≥ lim Kλ H ∞ (Dn ) = lim n

thus Kλ 

H ∞ (D∞ ) 2

=



s≥1 (1

n

n 

225

kλs H ∞ (D) =

s=1



(1 − |λ s |)−1 ,

s≥1

−1

− |λ s |) , and the result follows.



(b) Show with an example that there exists a function F ∈ H 2 (D∞ 2 ) such that ) where F (ζ) = F(rζ). for any r, 0 < r < 1, we have Fr  H ∞ (D∞ r 2 1 Solution: F = Kλ where λ ∈ D∞ 2 \ l , for example λ = (1/2, 1/3, 1/4, . . . ). Then, ∞ ∞  by (a), (Kλ )r = Krλ  H (D2 ).

(c) Show that U(H ∞ (D))  H ∞ (D∞ 2 ). Hint We admit the following profound theorem of Green and Tao (2008): the sequence P = (p j ) j≥1 of prime integers contains arbitrarily long (finite) arithmetic progressions. Solution: It is easy to see (verify!) that if the inclusion U(H ∞ (D)) ⊂ H ∞ (D∞ 2 ) ∞ ∞ ∞ were to hold, the mapping U : H (D) → H (D2 ) would be closed, hence bounded (closed graph theorem: Appendix E), and thus there would exist C > 0 such that U f H ∞ (D∞2 ) ≤ C f H ∞ (D) for every function f ∈ H ∞ (D). Show that the last majoration is impossible: let J ⊂ P = (p j ) j≥1 be a finite subset of primes and     f = j∈J a j z j , then U f H ∞ (D∞2 ) = sup|ζ j | 0, the function ζ → 1/(F(ζ)+ ) is in H ∞ (D∞ 2 ). By § 6.6.5(g), F/(F + ) ∈ E F . Since |F(ζ)/(F(ζ) + )| ≤ 1 and lim →0 (F(ζ)/(F(ζ) + )) = 1 for  every ζ ∈ D∞ 2 , we have (for the weak topology) 1 = lim →0 (F/(F + )) ∈ E F .

(b) Let F ∈ H 2 (Dn ) ⊂ H 2 (D∞ 2 ). The following assertions are equivalent. (i) F is Mζ -cyclic in H 2 (D∞ 2 ). α (ii) the function F is (ζ )α∈Zn+ -cyclic in H 2 (Dn ).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

RH, Dilations, and H 2 in the Hilbert Multi-disk

226

Solution: (ii) ⇒ (i) is evident. For the converse, observe that for every α ∈ Z+ (∞) \  Zn+ we have ζ α ⊥ H 2 (Dn ), and hence for every polynomial q = q(ζ) = α∈Z+ (∞) cα ζ α we have P(qF) = P(q)F where P is the orthogonal projection on H 2 (Dn ), and in  particular, P(q)(ζ) = α∈Zn+ cα ζ α . Moreover, if limk qk F−12 = 0 then limk P(qk F)−  12 = 0.

(c) Let f ∈ H 2 (D), α ∈ Z+ (∞) and F(ζ) = f (ζ α ). The following assertions are equivalent. (i) F is Mζ -cyclic in H 2 (D∞ 2 ). (ii) the function f is outer (thus, cyclic in H 2 (D)).  k In particular, a function f = k≥0 ak zn , where n ∈ N \ {1}, is (T n )-cyclic if and  only if the function ϕ = k≥0 ak zk is outer. Solution: The same reasoning as for (b), but replacing Zn+ by {αk : k ∈ Z+ }.

(d) By using the Theorem of § 6.6.6(a), show the following. (i) f = a1 z + a2 z2 + a3 z3 is (T n )-cyclic if and only if |a1 | ≥ |a2 | + |a3 |. (ii) f = a1 z + a2 z2 + a3 z3 + a4 z4 is (T n )-cyclic in H02 if and only if q(D) ∩ a3 D = ∅, where q(z) = a1 + a2 z + a4 z2 . In particular, the condition |a1 | ≥ |a2 | + |a3 | + |a4 | is sufficient, but not necessary (consider the case a3 = 0). (iii) f = a1 z + a2 z2 + a3 z3 + a4 z4 + a5 z5 is (T n )-cyclic if and only if q(D) ∩ (|a3 | + |a5 |)D = ∅, where q(z) = a1 + a2 z + a4 z2 . (iv) f = z(λ − z)N , where |λ| > 1 and N ∈ N, is (T n )-cyclic if N
λ > 0. (v) The polynomials p1 = a0 z + a1 z2 + a2 z4 + a3 z8 + a4 z16 and p2 = a0 z + a1 z12 + a2 z144 + a3 z1728 + a4 z20736

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press



6.8 Notes and Remarks

227

are (T n )-cyclic or not simultaneously (and if and only if the function  ϕ = 4k=0 ak zk is outer, i.e. has all its roots in C \ D). However – in general – this is not the case for the pair p1 , p3 , where p3 = a0 z + a1 z2 + a2 z3 + a3 z4 + a4 z5 .

6.7.6 A Function (Dn )-cyclic in L2 (0, 1) (Kozlov, 1950; Akhiezer, 1965) Let f be an odd 2-periodic function such that f (x) = 1, 0 < x < 1. Show that f is (Dn )-cyclic in L2 (0, 1). Solution: We have f =

k≥0

4 sin(2k + 1)πx, π(2k + 1)

hence after the change of basis in § 6.6.2 we obtain √

2 √2 2 2 α α 2k+1 f = z and U f = ζ λ , π(2k + 1) π α≥0 k≥0 where λ = (0, 1/p2 , 1/p3 , . . . ). Thus U f is Mζ -cyclic in H 2 (D∞ 2 ) by § 6.6.7(b).



6.8 Notes and Remarks The name “Riemann ζ function” (dominant in mathematics) is questionable.  Indeed, the author of the definition of the function ζ(s) = n≥1 (1/n s ), as well as of its fundamental properties (its multiplicative representation in § 6.1.2 and the functional equation of Theorem 6.1.5) is known: it was Leonhard Euler. A presentation by Gelfond (1958) given in a colloquium devoted to Euler shows in a few lines that Euler’s computation in his 1761 note is equivalent (for s real) to the functional equation rediscovered by Riemann in 1859. The fact that Riemann had extended the definition to the plane C and thereby found profound links with the prime numbers, adds nothing to the question of the discovery. For example, nobody would dream of attributing the invention of the airplane to Willy Messerschmitt (1944) rather than to the brothers Wilbur and Orville Wright (1903) under the pretext that jet planes – whose first series production was the work of Messerschmitt – now dominate the sky; or, to remain within the subject of this book, the Fourier transform does not cease to be Fourier’s simply because in the framework of the space H 2 (C+ ) it is essentially complex. The name Riemann ζ function is apparently due to Helge von Koch

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

228

RH, Dilations, and H 2 in the Hilbert Multi-disk

(1902), “Ueber die Riemann’sche Primzahlfunction” (according to the website “Earliest Known Uses of Some of the Words of Mathematics,” http://jeff560 .tripod.com/mathword.html). One can hope that this historical error will soon be corrected. For very convincing mathematical and historical arguments for the same point of view see also Blagouchine (2018) and the references therein. The properties in § 6.1.1–6.1.2 and Lemma 6.1.3 form the standard basis of the theory of the function ζ: the product in § 6.1.2 is due to Euler (1737, published in 1743), and the integral representations of ζ(s) originate with Riemann (1859). Today numerous integral formulas are known (the web page http://functions.wolfram.com/Zeta contains 198 representative formulas for zeta). The functional equation is implicitly presented by Euler, especially for integer s, but also for certain rationals (1761: see the explanations of Gelfond (1958, pp. 89–90), mentioned above), and then reappears with Riemann (1859), in the complex domain and with two different proofs. Today several proofs are known; that of § 6.1.5 is a combination of a proof by Titchmarsh (1951, Ch. II, § 1) and one by Hardy (1922). In total, Titchmarsh (1951) gives seven different proofs of the functional equation. The Riemann hypothesis (§ 6.1.3) is the most celebrated unsolved problem in mathematics. It was part of Hilbert’s famous list of 23 problems for the twentieth century (presented in his speech at the 1900 International Congress of Mathematicians in Paris); more precisely, problem no. 8 of the list mentions the Riemann hypothesis, the Goldbach conjecture, and the twin prime problem. According to an anecdote, Hilbert jokingly stated that if 500 years after his death he had the right to return to this world for 30 seconds, he would use it to pose the question “Has RH been resolved?” The RH provoked an enormous amount of activity in mathematics throughout the twentieth century, and in the year 2000 was included in the list of seven “Millennium Prize Problems” (problem no. 4 of the list; the resolution of each of the problems is rewarded with a prize of a million US dollars offered by the Clay Mathematics Institute (Cambridge, USA)). The literature on the problem is immense; the classical sources remain the books by Titchmarsh (1951), Landau (1927), and Hardy and Wright (1938). For modern surveys see, for example, the official presentation for the Millennium Prize by Enrico Bombieri (2000) and a summary of the latest advances by Peter Sarnak (2005); both are available on the website of the Clay Institute, www.claymath.org/millenniumproblems/riemann-hypothesis. See also Conrey (2003), a synthesis article by P´erez-Marco (2011), and for a “light literary” history of the RH, Sabbagh (2002). The approach to the RH by approximation, presented in § 6.2–6.5, is one of the dozens of equivalent forms of the RH; several of them are collected on

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.8 Notes and Remarks

229

the website of the American Institute of Mathematics (Palo Alto, California), www.aimath.org/WWN/rh/; for more, see P´erez-Marco (2011). Among the more curious conjectures equivalent to the RH are 1

√

2

n+ n log n for every n ≥ 74 (Schoenfeld, 1976); – LCM(1, 2, 3, . . . , n) ≤ e 8π  2 −1+

), ∀ > 0, as N → ∞; here 1/N = b1 < b2 < – 1≤ j≤A (b j − j/A) = O(N · · · < bA = 1 are the fractions of John Farey (1816, an English geologist) defined by {b1 < b2 < · · · < bA } = {h/k : 0 < h ≤ k ≤ N} (Franel, 1924).

These last two forms of the RH, as well as commentaries, references, and 17 further equivalent forms can be found in a survey by Balazard (2010). The approach of § 6.2–6.5 was proposed by Nyman (1950). (Note that Bertil Nyman remains a very enigmatic figure: after a brilliant thesis, he completely disappeared from the mathematical world.) Several other conjectures equivalent to RH are expressed in the language of approximations. For example, Norbert Wiener (him again!) mentioned in his famous work on the Tauberian theorems (Wiener, 1932) that for every σ, 0 < σ < 1, the completeness of the translations (τ s fσ ) s∈R in the space L1 (R), where  e x  fσ (x) = e(σ−1)x ex , e −1 is equivalent to the fact that, for every t ∈ R, ζ(σ + it)  0. Then, Salem (1953) showed that this last property is also equivalent to the completeness of the dilations (Dt gσ )t>0 in L1 (R+ ), where gσ (x) = xσ−1 (e x + 1)−1 . Levinson (1956) provided a similar criterion for the absence of zeros of ζ(s) for 1/2 ≤ σ1 < Re(s) < σ2 ≤ 1. The form of Nyman’s criterion is particularly advantageous because it introduces the ability to use the classical techniques of the space L2 on a compact interval. Moreover, the criterion of Theorem 6.4.1 was considerably reinforced by B´aez-Duarte (2003), as is mentioned in § 6.6, bringing it closer to the Wintner–Beurling problem treated in § 6.6.1–6.6.7. Part (2) of Theorem 6.2.1 is taken from Nyman (1950); the quantitative part (1) is a simple clarification of Nyman’s reasoning. It is nonetheless interesting to compare it numerically (as well as with the estimations in § 6.5) with the best bounds previously known, obtained by a “heavy artillery” approach due to Korobov and Vinogradov (1958; see Ford (2002) for a modern presentation), specifically with the fact that ζ(s)  0 in the domain 

1 , |y| ≥ 3. Ω = s = x + iy : x = Re(s) ≥ 1 − (57.54)(log |y|)2/3 (log log |y|)1/3

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

230

RH, Dilations, and H 2 in the Hilbert Multi-disk

It seems that at the level y = 3 the disk 1  D 2,r , d

2

1 −1 d2 √ of Theorem 6.2.1(1) is better than Ω if 0 < d < 1/ 10 ≈ 1/3. The classical Paley–Wiener theorem § 6.3.2–6.3.3(a) is presented following Nikolski (1980, 1986), but under a “weak form,” hence without a description of the spaces H 2 (C+ ), H 2 (C+ ) by quadratic means: ⎧ ⎫  ⎪ ⎪ ⎨ ⎬ 2 2 H (C+ ) = ⎪ ⎩ f ∈ Hol(C+ ) : sup | f (x + iy)| dx < ∞⎪ ⎭. 1 r= d

y>0

R

As is well known, to reach such a description we need a kind of embedding theorem, for example that of Gabriel and Zygmund mentioned in § 2.9. The assertions of § 6.3.2(c), due to Peter Lax, are heavily used in diffusion theory: see Lax and Phillips (1967). The corollaries of § 6.3.4 touch on a very important theme in analysis, with several applications to harmonic analysis, signal processing, and stochastic processes; see, for example, Nikolski (2002) and Rozanov (1963). We must point out that, as in the case of the circle T (see Theorems 2.7.4 and 2.7.5), there exists in addition the question of the completeness of the exponentials, the translations, and the dilations in a corresponding “bilateral” space (L2 (R) or L2 (R+ , dx/x)). For example, in the same manner as in Chapter 2, using the results of this chapter we obtain: (i) the translations (τ s f ) s∈R generate the space L2 (R) if and only if F f  0 a.e. on R; (ii) the translations (τ s f ) s∈R+ generate the space L2 (R) if and only if F f  0 a.e. on R and  log |F f | dx = −∞ 2 R 1+x (Paley and Wiener, 1934). Section 6.5 follows Nikolski (1995). Section 6.6 is an excerpt from Nikolski (2012). The problem of the completeness of the integer dilations (Dn f )n≥1 appears naturally in view of the theorem of B´aez-Duarte (2003) cited in § 6.6. We mention that Bagchi (2006) further simplified the form of this last criterion for RH by reducing it to the following proposition. Theorem (Bagchi, 2006) Let l2 (1/n2 ) be the weighted space



|xn |2 n−2 < ∞ , l2 (1/n2 ) = x = (xn )n≥1 : n≥1

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

6.8 Notes and Remarks

231

and xl = (ρ(n/l))n≥1 (ρ(x) = x − [x]), xl ∈ l2 (1/n2 ), l = 2, 3, . . . The following assertions are equivalent. (1) (2) (3) (4)

The RH is correct.   spanl2 (1/n2 ) xl : l = 2, 3, . . . = l2 (1/n2 ).   1 = (1, 1, . . . ) ∈ spanl2 (1/n2 ) xl : l = 2, 3, . . . . The same as (2) and/or (3) but for xl with l not containing any squares.

For its proof, this even more elementary form of Nyman’s theorem requires nonetheless additional techniques of the theory of the function ζ: a decomposition in a Dirichlet series of 1/ζ(s), and then the theorems of Lindel¨of, Littlewood, the functional equation of ζ, etc. In reality, the problem of completeness of the integer dilations (Dn f )n≥1 appeared long before these results, namely in Wintner (1944), motivated by certain problems in Diophantine analysis and where the first profound results were obtained. Independently, Beurling (1945) presented this problem to a seminar in Uppsala. The importance of the Wintner–Beurling problem was widely recognized in the 1950s (see, for example, the important publications of Bourgin (1946) and Kozlov (1948, 1950) who, of course, had no knowledge of Beurling’s seminar) but virtually forgotten for another 40 years. Concerning the renewal of interest in the completeness of the dilations (which can, in fine, help clarify the RH), see Hedenmalm et al. (1997, 1999) and references therein. The Bohr transform, § 6.6.3, was introduced in Bohr (1913), but for a study  of the Dirichlet series n≥1 (an /n s ) (always linked to the cluster of ideas around the RH). The article of Hilbert (1909) mentioned in § 6.6.4 was only a research plan that subsequently was (partially) successful. Hilbert strongly insisted on the absolute convergence of the power series of an infinite number of variables, without which the subject becomes quite fuzzy. Corollary 6.6.3 is the principal result of Beurling’s presentation in Uppsala (Beurling, 1945), which was proved without passing to the space H 2 (D∞ 2 ). The last remark also applies to the result § 6.6.6(a) (Neuwirth et al., 1970). The lemma § 6.6.6(b) is a somewhat improved form of a result from the last article (which was later rediscovered by several authors). Corollary § 6.6.7(c) is from the founding article by Wintner (1944) (with a different proof), where Wintner was motivated by problems linked to the Sieve of Eratosthenes (hence, by the distribution of prime numbers). Most of the propositions of § 6.7 are borrowed from Nikolski (2012). The example §6.7.6 is a special case of the results of Kozlov (1950), where the following problem was posed. Let θ, 0 < θ ≤ 1, and let fθ be an odd 2-periodic function such that for 0 < x < 1, fθ (x) = χ(0,θ) (x); give a criterion of (Dn )-

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

232

RH, Dilations, and H 2 in the Hilbert Multi-disk

cyclicity of fθ (as a function of θ). Kozlov stated that fθ is cyclic for θ = 1 (this is §6.7.6; in fact, this theorem was proved in Akhiezer (1965, Section “Additions and Problems,” §I.23), but with a (fairly long) proof completely different to ours), θ = 1/2 and θ = 2/3, and is not cyclic for θ = 1/3 (and for all the θ in a neighborhood of 1/3), as well as for θ admitting a representation θ = q/p where p > 2 is a prime and q odd such that tan2 (qπ/2p) < 1/p (but this condition is not satisfied for q = 1, p = 3, which corresponds to θ = 1/3).

https://doi.org/10.1017/9781316882108.008 Published online by Cambridge University Press

Appendix A Key Notions of Integration

A.1 Measures Let Ω be a set, and A a σ-algebra on Ω, so that (Ω, A) is a measurable space. A positive measure on A (or on Ω) is a countably additive mapping μ : A →    R+ = R+ {∞}, such that μ( A j ) = j μ(A j ) for every disjoint sequence (A j )  (A j Ak = ∅, j  k). A triplet (Ω, A, μ) is called a measure space. A complex measure is a countably additive mapping μ : A → C. The set of complex measures is denoted M(Ω) (when the σ-algebra is evident). function μ : A → R+ which is additive and countably sub-additive (i.e. • A  μ( A j ) ≤ j μ(A j ) for every sequence (A j )) is a measure. supA∈A |μ(A)| < ∞, and is of finite • A complex measure μ is always bounded,  total variation μ = Var(μ) := sup j |μ(A j )| where the sup is taken over all disjoint finite families (A j ). The variation of μ is a measure |μ| defined by |μ|(A) = μ|A, A ∈ A. If μ ≥ 0 then |μ| = μ. • A real measure μ admits a unique representation of the form μ = μ1 − μ2 , where μ j ≥ 0 and there exists A ∈ A such that μ1 (A) = 0, μ2 (Ω\A) = 0. Moreover, |μ| = μ1 + μ2 . • A complex measure μ admits a representation of the form μ = μ1 − μ2 + iμ3 − iμ4 , where μ j ≥ 0. • The set of positive measures on Ω is a lattice: for every sequence (μn ), μn ≥ 0, there exists a unique μ = supn μn such that for every n, μn ≤ μ, and if μn ≤ ν for every n, then μ ≤ ν. The measure μ is given by

μk Ak , μA = sup k≥1

233 https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

234

Key Notions of Integration

where the sup is taken over all disjoint partitions A = k  j). Similarly, there exists ν = inf n μn , with

μk Ak . νA = inf

 k

Ak (Ak



A j = ∅,

k≥1

• Let P(x) be a property defined for x ∈ Ω; P is said to hold μ-almost everywhere (abbreviated as μ-a.e.) if μ(x ∈ Ω : P(x) does not hold) = 0. • Measures on a topological space. A topological space Ω, by default, is equipped with the Borel σ-algebra, A = B: this is the σ-algebra generated by the open subsets (or, likewise, by the closed subsets) of Ω. A measure on B is said to be a Borel measure. It is called regular if ∀B ∈ B, ∀ > 0, ∃A closed, ∃C open, such that A ⊂ B ⊂ C and μ(C \ A) < . • If the space Ω is locally compact and σ-compact and if μ is a Borel measure which is finite on every compact subset, then μ is regular; in particular, any locally finite measure in Rn is regular. supp(μ) of a Borel measure on Ω is defined by • The (closed) support  Ω \ supp(μ) = O (O runs over all the open subsets such that μ(O) = 0).

A.2 The Lebesgue Integral Let (Ω, A, μ) be a measure space. A function f : Ω → C is said to be measurable if f −1 (B) ∈ A for every Borel set B ⊂ C (or, equivalently, for every rectangle (a product of intervals) B ⊂ R2 = C). The integral of a positive measurable function is  n

 f dμ := sup ck μAk : 0 ≤ ck ≤ f (x), x ∈ Ak ; Ak ∩ A j = ∅(k  j) , Ω

k=1



f is integrable if Ω f dμ < ∞. A function f : Ω → R is integrable if f ± := max(0, ± f ) are integrable, and we set    f dμ = f + dμ − f − dμ. Ω

Ω

Ω

A function f : Ω → C is integrable if Re( f ) and Im( f ) are integrable, and we   set Ω f dμ = Ω Re( f ) dμ + i Ω Im( f ) dμ.

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

A.2 The Lebesgue Integral

235

Henri Lebesgue (1875–1941) was a French mathematician and creator of the modern theory of integration, which changed the face of mathematics. With origins in a modest provincial background, he followed (thanks to the efforts of his mother) the complete cycle of French education, including the preparatory classes at the Lyc´ee Louis-le-Grand in Paris, ´ and then the Ecole Normale Sup´erieure (1897). Upon graduation, he obtained only a modest position as a high school teacher (in the Lyc´ee Central in Nancy), in 1899–1902, when he wrote his famous article “Sur une g´en´eralisation de l’int´egrale d´efinie” (Comptes rendus de l’Acad´emie des sciences (1902)) and prepared his thesis “Int´egrale, longueur, aire” (130 pages) submitted in Paris in 1902. Measure and integration theory were thus created, stimulating an explosion of developments in harmonic analysis and in mathematics in general. Even though his work met a fairly hostile reception in France (he had a prolonged rivalry with Baire, and only obtained his first university position in Paris, maˆıtre de conf´erences at the Sorbonne, in 1910), his new theory rapidly gained ground internationally. After 10–15 years, those areas of mathematics that required the Lebesgue integral (with the enthusiastic participation of Hardy, Littlewood, Frigyes and Marcel Riesz, Hausdorff, Steinhaus, Borel, Denjoy, Fatou, Nikodym, Banach, Plancherel, Luzin, Kolmogorov, Radon, Saks, Haar, etc.) had changed beyond all recognition. Lebesgue published two important monographs on the subject: Lec¸ons sur l’int´egration et la recherche des fonctions primitives (1904) and Lec¸ons sur les s´eries trigonom´etriques (1906). But he was not content to limit himself to the pure theory of integration – after all, he was the author of the famous saying: R´eduites aux th´eories

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

236

Key Notions of Integration

g´en´erales, les math´ematiques deviendraient une belle forme sans contenu (“Reduced to general theories, mathematics would become a beautiful framework without content”). Lebesgue made a variety of important contributions: in topology, in potential theory, on the Dirichlet problem, in the calculus of variations, in set theory, and in the theory of dimensions. In 1922 he published a summary of his 90 articles and books, Notice sur les travaux scientifiques de M. Henri Lebesgue – a work of synthesis with certain evaluations of his major results. Paul Montel described his final days: “At the beginning of 1941, Henri Lebesgue gave his last annual course at the Coll`ege de France. Already, the sickness that took him a few months later added to the low morale caused by the defeat and the enemy occupation. He could barely walk, and the city was severely lacking in surface transport. In order to give his lectures, he had to rely on the wheelchairs and bicycles that were used to transport the sick.” Lebesgue was elected member of several Academies: l’Acad´emie des Sciences (Paris), the Royal Society, the Belgian Acad´emie Royale, the Academy of Bologna, the Accademia dei Lincei (Rome), the Romanian Academy of Science, and the Krakow Academy of Sciences. 1 • The set of integrable functions L (μ) is a vector space and the integral  f −→ f dμ Ω

is a linear functional on L1 (μ) satisfying      ≤ f dμ | f | dμ.   Ω

Ω

If f, g ∈ L1 (μ) are real and f (x) ≤ g(x) μ-a.e., then   f dμ ≤ g dμ. Ω

Ω

• If μ ∈ M(Ω) with the decomposition μ = μ1 − μ2 + iμ3 − iμ4 , where μ j ≥ 0 (see above), and if f ∈ L1 (|μ|), then we set      f dμ = f dμ1 − f dμ2 + i f dμ3 − i f dμ4 . Ω

Ω

Ω

Ω

• Passage to the limit.

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

Ω

A.2 The Lebesgue Integral

237

(1) The Beppo Levi Theorem. If fn (x)  f (x) and fn (x) ≥ 0 μ-a.e., then   lim fn dμ = f dμ. n

Ω

Ω

 The same holds for fn  f if we suppose that Ω f1 dμ < ∞. (2) The Lebesgue dominated convergence theorem. If limn fn (x) exists for almost all x ∈ Ω, and for every n, | fn | ≤ f μ-a.e. with f ∈ L1 (μ), then   fn dμ = (lim fn ) dμ. lim n

Ω

Ω

n

(3) Fatou’s lemma. If fn ≥ 0, then   (lim fn ) dμ ≤ lim fn dμ. Ω

n

n

Ω

• Integrals depending on a parameter. Let K be a metric space, and let f : Ω × K → C be a mapping such that, for all t ∈ K, f (·, t) ∈ L1 (μ), and  f (x, t) dμ(x), t ∈ K. F(t) = Ω

(1) Continuity. If there exists a function h ∈ L1 (μ) such that ∀t ∈ K, | f (x, t)| ≤ h(x) μ-a.e. and limt→t0 f (x, t) = f (x, t0 ) μ-a.e., then F is continuous at the point t0 . (2) Differentiability. Let K ⊂ R be an open set, and suppose that for every (x, t) ∈ Ω × K there exist functions g(x, t) :=

∂ f (x, t) ∂t

and h ∈ L1 (μ) such that ∀t ∈ K, |g(x, t)| ≤ h(x) μ-a.e. Then F is  differentiable on K, and F  (t) = Ω g(x, t) dμ(x). (3) Holomorphy. Let K be an open subset of C, t −→ f (x, t) holomorphic in K and | f (x, t)| ≤ h(x) for every (x, t) ∈ Ω × K where h ∈ L1 (μ). Then F is holomorphic on K. • Primitive of a integrable function and Lebesgue points. Let I ⊂ R be an interval and f ∈ L1 (I, dx). h (i) For almost every point x ∈ I, limh→0 h−1 0 | f (x) − f (x + t)| dt = 0 (such an x is called a Lebesgue point of f ). h (ii) At every Lebesgue point x, f (x) = limh→0 h−1 0 f (x + t) dt.

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

238

Key Notions of Integration

A.3 Lebesgue Decomposition and the Radon–Nikodym Theorem Let (Ω, A, ν) be a measure space. • Lebesgue decomposition. Let μ ∈ M(Ω). There exists a unique decomposition μ = μa + μ s where μa , μ s are two measures such that: (i) ∀A ∈ A, ν(A) = 0 ⇒ μa (A) = 0; μa is said to be absolutely continuous with respect to ν; this is denoted by μa  ν. (ii) μ s ⊥ ν in the sense that there exists A ∈ A such that |μ s |(A) = 0, ν(Ω \ A) = 0 (μ s is said to be singular with respect to ν). In fact, μa = χA μ, μ s = (1 − χA )μ. • The Radon–Nikodym theorem. Let μ ∈ M(Ω). Then μ  ν ⇔ there exists a function h ∈ L1 (μ) such that μ(A) = A h dν (∀A ∈ A); μ is called a measure with density   h and can be written μ = hν (or h = dμ/dν). We have f dμ = f h dν for every f ∈ L1 (|μ|). Ω Ω We always have μ  |μ| and = dμ/d|μ| is unimodular |μ|-a.e.

A.4 The Riesz Representation Theorem Let Ω be a compact space and C(Ω) the space of continuous functions on Ω equipped with the uniform norm  f ∞ = supΩ | f |, and let ϕ be a linear functional on C(Ω). The following assertions are equivalent. (1) ϕ is continuous (bounded). (2) There exists a complex measure μ ∈ M(Ω) such that, for every function  f ∈ C(Ω), ϕ( f ) = Ω f dμ. Such a measure μ is unique, and ϕ = μ. Note that Frigyes Riesz proved the theorem for Ω = [0, 1] (1909), Banach for metric spaces Ω (1933), and Kakutani for the general case (1941); see Rudin (1998) for comments. • For a locally compact space Ω, the same statement holds for C0 (Ω) = { f ∈ C(Ω) : ∀ > 0, ∃K a compact set such that| f | < on Ω \ K}.

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

A.5 The Lebesgue L p (μ) Spaces

239

A.5 The Lebesgue L p(μ) Spaces Let (Ω, A, μ) be a measure space and 0 < p < ∞. We define 

 p p L (μ) = L (Ω, μ) := f : Ω → C measurable : | f | p dμ := N p ( f ) p < ∞ . Ω

L (μ) is a vector space and, if p ≥ 1, N p is a seminorm on L p (μ). p

p p p • The Lebesgue space is the normed space L (μ) = L (μ)/R of L (μ) modulo the equivalence relation R( f, g) ⇔ f = g μ-a.e. It is a complete normed space, and hence a Banach space, equipped with the norm F p = N p ( f ), ∀ f ∈ F ∈ L p (μ). For p = ∞,

 L∞ (μ) := f : Ω → C measurable : N∞ ( f ) = inf{λ > 0 : μ(| f | > λ) = 0} < ∞ . p • The distribution function and weak L spaces. Let f : Ω → C be a measurable function and

λ f (t) = μ(x ∈ Ω : | f (x)| ≥ t), t > 0. ∞ Then, N p ( f ) p = p 0 t p−1 λ f (t) dt, and if f ∈ L p (μ), then λ f (t) = o(t−p ) when t → ∞. The space L p,∞ (“L p weak”) is defined as the set of functions f such that λ f (t) = o(t−p ) when t → ∞. p • The L spaces and the Lebesgue decomposition (Radon–Nikodym). Let μ = μa + μ s , where μa = χA μ and μ s = (1 − χA )μ, be the Lebesgue decomposition (see above). For a function f ∈ L p (μ), by setting fa = χA f , f s = (1 − χA ) f , we obtain fa ∈ L p (μa ), f s ∈ L p (μ s ) and f = fa + f s ,  f Lp p (μ) =  fa Lp p (μa ) +  f s Lp p (μs ) . This is clearly a direct decomposition, denoted L p (μ) = L p (μa ) ⊕ L p (μ s ) (direct sum of type l p ). p • H¨older’s inequality. If f j ∈ L j (μ), p j > 0 ( j = 1, . . . , n), and

1 1 = , s pj 1 n

then

n 1

f j ∈ L s (μ) and  n 1

n  f j ≤  f jp j . s

1

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

240

Key Notions of Integration

The classical special case: for n = 2, 1 = 1/p + 1/p , then 

f ∈ L p (μ), g ∈ L p (μ) ⇒  f g1 ≤  f  p g p . For p = 2, this becomes the Cauchy–Schwarz inequality:  f g1 ≤  f 2 g2 . • The converse of H¨older’s inequality. Let f be a measurable function and 1 ≤ p ≤ ∞, then 

f ∈ L p (μ) ⇔ ( f g ∈ L1 (μ), ∀g ∈ L p (μ));

 moreover, sup{| Ω f g dμ| : g p ≤ 1} =  f  p . • Jensen’s convexity inequality. If ϕ is a convex function defined on an interval I ⊂ R where a real function f takes its values ( f (Ω) ⊂ I), then for every positive finite measure μ,     1 1 f dμ ≤ ϕ ◦ f dμ. ϕ μ(Ω) Ω μ(Ω) Ω (This follows from the fact that ϕ(x) = sup{L(x) : L linear and L ≤ ϕ} and for linear L, the inequality is a trivial equality.) Borel measure with compact • Density of the polynomials. Let μ be a finite  support in Rn . Then the polynomials f (x) = α≥0 aα xα , xα = x1α1 x2α2 . . . xnαn , α = (α1 , . . . , αn ) ∈ Zn+ , are dense in L p (μ), p < ∞. Outline of a direct proof: (a) the polynomials are dense in the space C = C(supp(μ)) (theorem of (Stone–)Weierstrass), hence it only remains to show that closL p (μ) (C) = L p (μ); (b) we show that for any compact set F, χF ∈ closL p (μ) (C) (χF = limn fn where fn (x) = (1 − min(dist(x, F), 1))n ); then (c) for every A ∈ B, χA ∈ closL p (μ) (C) (by the regularity of μ); finally (d)  L p (μ) = closL p (μ) (C).

A.6 Convolution and the Fourier Transform If G is a locally compact commutative group (such as Tn , Zn , Rn ), the convolution of two measures μ, ν ∈ M(G) is defined using the Riesz representation theorem as the measure μ ∗ ν such that, for every ϕ ∈ C0 (G),    ϕ d(μ ∗ ν) = ϕ(x + y) dμ(x) dν(y). G

G

G

• We have μ ∗ ν = ν ∗ μ, μ ∗ ν ≤ μ · ν, hence M(G) is a commutative Banach algebra (see Appendix D) with unit δ0 , the Dirac delta at the origin.

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

A.6 Convolution and the Fourier Transform

241

• The Fourier transform. Let Gˆ be the dual group of the unimodular continuous multiplicative characters of G and μ ∈ M(G); the Fourier transform of μ is defined by  ˆ γ(−x) dμ(x), γ ∈ G. F μ(γ) = G

Remark For a reason of normalization (linked especially to Plancherel’s theorem, see below), when we apply the definition to a measure μ = f m absolutely continuous with respect to the invariant measure m (Haar measure), we use an embedding L1 (m) ⊂ M(G), f −→ c f m, selecting a constant c in order to have F (F f ) = f for certain test functions. In particular, c = (2π)−n/2 in the case of G = Rn , c = (2π)−n in the case of Tn = Rn /Zn , c = 1 in the case of Zn , hence  1 f (x)e−ixt dx, f ∈ L1 (R), t ∈ R, F f (t) = √ 2π R 1 F f (n) = fˆ(n) = 2π

 R/Z

f (x)e−ixn dx,

f ∈ L1 (T), n ∈ Z.

• For every μ ∈ M(G), F μ is bounded and uniformly continuous; for f ∈ L1 (m), limγ→∞ F f (γ) = 0 (the Riemann–Lebesgue lemma, correct for (at least) the classical groups Tn , Rn ). ˆ formula. For every μ ∈ M(G) and ν ∈ M(G), • Transfer   F μ dν = G F ν dμ. This follows from Fubini’s theorem. Gˆ n • Uniqueness theorem. F μ = 0 ⇒ μ = 0. (In the case G = T , this follows from the preceding formula and Weierstrass’s theorem.)

ˆ F (μ ∗ ν)(γ) = F μ(γ)F ν(γ). • For every μ, ν ∈ M(G) and every γ ∈ G, • The Fourier–Plancherel transform. With a proper normalization (mentioned above), F : (L1 (G) ∩ L2 (G)) → L2 (G) is an isometric mapping with a dense image, hence it can be extended in a unique manner to a unitary operator F : L2 (G) → L2 (G) 4

such that F (F f )(x) = f (−x), ∀ f , and hence F = id. For every f ∈ L2 (G), limK F f − F ( f χK )2 = 0, where K runs over the compact subsets “filling G” (for example, in the case of R, K running over the intervals [−t, t], t > 0).

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

242

Key Notions of Integration

p p • Convolution in L (G).  Let f ∈ L (G) and τ s f (x) = f (x − s), s ∈ G; then the convolution f ∗ μ = G τ s f dμ(s) is well-defined in L p (G), and  f ∗ μ p ≤  f  p μ. • Approximate identities, Fej´er polynomials.

(i) If (μk )k≥1 are measures on T (or Tn ) such that supk μk  < ∞, and if for every n ∈ Z, limk μˆ k (n) = 1, then for every f ∈ L p (T), 1 ≤ p < ∞, limk  f − f ∗ μk  p = 0. (This follows from the density of the trigonometric polynomials in L p (T).) (ii) In particular, for μk = Φk m, where

(1 − | j|/k)ei jx = k−1 (sin(kx/2)/ sin(x/2))2 Φk (eix ) = | j|≤k

(Fej´er kernel), limk  f − f ∗ Φk  p = 0 (∀ f ∈ L p (T)).

https://doi.org/10.1017/9781316882108.009 Published online by Cambridge University Press

Appendix B Key Notions of Complex Analysis

B.1 Analytic Functions and Holomorphic Functions Let Ω be an open subset of the complex plane C and f : Ω → C a function in Ω. The following assertions are equivalent. (1) f is analytic in Ω: ∀z ∈ Ω, ∃r > 0 such that D(z, r) ⊂ Ω and for every  ζ ∈ D(z, r) f (ζ) = k≥0 ak (ζ − z)k (absolute convergence ). (2) f is holomorphic in Ω: f ∈ C 1 (Ω) and ∂f ∂z where ∂ ∂z

=

= 0 in Ω,

∂ 1 ∂ +i , 2 ∂x ∂y

z = x + iy ∈ Ω.

This equation is called the Cauchy–Riemann (C-R) equation. In particular, a holomorphic function f is in C ∞ (Ω); its derivative ∂ f /∂z = 12 (∂ f /∂x − i(∂ f /∂y)) is denoted f  (z) (complex derivative of f ). By separating the real part u = Re( f ) and the imaginary part v = Im( f ), we obtain another form of the C-R equation: ∂v ∂u ∂u ∂v = , =− . ∂x ∂y ∂x ∂y • The set Hol(Ω) of holomorphic functions in Ω is a vector space.

B.2 Harmonic Functions, Forms, and Primitives A function u ∈ C 2 (Ω) is said to be harmonic if Δu = 0 in Ω, where Δ = ∂2 /∂x2 + ∂2 /∂y2 is the Laplacian operator. The set of harmonic functions on Ω 243 https://doi.org/10.1017/9781316882108.010 Published online by Cambridge University Press

244

Key Notions of Complex Analysis

is a vector space. Given C-R, the real and imaginary parts Re( f ) and Im( f ) of a holomorphic function f are harmonic, and hence so is f . • Two real harmonic functions u, v are called harmonic conjugates if there exists a function f ∈ Hol(Ω) such that u = Re( f ), v = Im( f ) (or, equivalently, u, v satisfy the C-R system). 1 • Recall that a differential form α = Pdx + Qdy (where P, Q ∈ C (Ω)) is said to be closed if dα = 0, where dα = (∂P/∂y − ∂Q/∂x) dx ∧ dy, and exact if there exists a primitive v ∈ C 2 (Ω) of α, i.e. v such that dv = α, where dv = (∂v/∂x) dx + (∂v/∂y) dy. An exact form is always closed.

• For an open subset Ω of the complex plane C the following assertions are equivalent. (1) Every closed form in Ω is exact. (2) Every real harmonic function in Ω admits a harmonic conjugate. (3) Every holomorphic function in Ω admits a holomorphic primitive: f ∈ Hol(Ω) ⇒ ∃F ∈ Hol(Ω) such that F  (z) = f (z), z ∈ Ω. (4) Ω is simply connected (i.e. every continuous closed curve is homotopic to a point: “there are no holes in Ω”). Remark (2) follows from (1) by applying it to α = (∂u/∂y) dx − (∂u/∂x) dy where u is harmonic (Δu = 0). The standard example of a harmonic function without a conjugate is u(z) = log |z|, z ∈ Ω = C \ {0}, and that of a holomorphic function without a primitive is f (z) = 1/z, z ∈ Ω = C \ {0}.

B.3 Integral Formulas If f ∈ Hol(Ω) and if γ is a closed curve homotopic to a point in Ω, then f (z) dz = 0 (a form α = f (z) dz is closed). γ • If f ∈ Hol(Ω), Ω is simply connected, and γ is a simple closed curve in Ω, then for every ζ ∈ int(γ),  1 f (z) dz f (ζ) = 2πi γ z − ζ (Cauchy’s formula). • If u is harmonic in Ω and D(ζ, r) ⊂ Ω (r > 0), then    1 1 u(ζ) = u(z)|dz| = 2 u(x + iy) dx dy 2πr ∂D(ζ,r) πr D(ζ,r) (mean-value formulas).

https://doi.org/10.1017/9781316882108.010 Published online by Cambridge University Press

B.4 Major Principles of Complex Analysis

245

B.4 Major Principles of Complex Analysis Let Ω be an open subset of C. • Principle of isolated zeros. If Ω is connected, then for every function f ∈ Hol(Ω), f  0 and any ζ ∈ Ω there exists r > 0 such that D(ζ, r) ⊂ Ω and f (z)  0 for every 0 < |z − ζ| < r. Consequently, the set of zeros of f is either finite, or else form a sequence tending to the boundary ∂Ω. Remark Let f, g ∈ Hol(Ω) and let γ be a simple continuous closed curve in a simply connected Ω; let N( f, γ) denote the number of zeros of f in int(γ). We cite Rouch´e’s theorem for the zeros of a “perturbed” function: N( f + g, γ) = N( f, γ) if |g(z)| < | f (z)| for z ∈ γ. • Maximum principle. If Ω is bounded, then for every harmonic function u (in particular, for every holomorphic function) and any ζ ∈ Ω,   |u(ζ)| ≤ sup lim |u(z)| , λ∈∂Ω z→λ,z∈Ω

and equality holds only if u = constant on the connected component of Ω containing ζ. • The compactness principle (Montel). If there is a sequence ( fn ) ⊂ Hol(Ω), uniformly bounded on every compact subset K of Ω, that is,  fn C(K) = sup | fn (z)| ≤ cK < ∞ for every n = 1, 2, . . . , z∈K

then there exists a subsequence ( fn j ) converging uniformly on any compact subset K ⊂ Ω to a function f ∈ Hol(Ω) : lim j  f − fn C(K) = 0. Principle of conformal mappings (Riemann). Every connected and simply connected open set Ω ⊂ C, Ω  C is conformally equivalent to the unit disk D = D(0, 1) (and hence they are all conformally equivalent to each other): there exists a bijective and biholomorphic mapping (said to be conformal) ϕ : Ω → D, ϕ ∈ Hol(Ω), ϕ−1 ∈ Hol(D). Remarks (1) A Jordan domain Ω is a bounded open set whose boundary ∂Ω is homeomorphic to the unit circle T (⇔ it is a simple, continuous, and closed curve); every conformal mapping ϕ : D → Ω on a Jordan domain can be extended to a homeomorphism of D onto Ω (and in particular, ϕ ∈ C(D)) (Carath´eodory, 1913). (2) Every conformal mapping ϕ of D onto itself is of the form ϕ(z) =

z−λ 1 − λz

https://doi.org/10.1017/9781316882108.010 Published online by Cambridge University Press

246

Key Notions of Complex Analysis where | | = 1, |λ| < 1. Every conformal mapping of C+ = {z : Im(z) > 0} on D is of the form z−λ ϕ(z) =

z−λ where | | = 1, λ ∈ C+ .

B.5 Holomorphic Extensions Let f ∈ Hol(Ω) and λ ∈ ∂Ω. The function f is said to be (holomorphically) extendable at a point λ if there exists r > 0 and g ∈ Hol(D(λ, r)) such that f = g on Ω ∩ D(λ, r). • A function f ∈ Hol(Ω) is extendable at the point λ ∈ ∂Ω if and only if there exists ζ ∈ Ω such that the radius of convergence R of the local development  f (z) = k≥0 ak (z − ζ)k satisfies R > |ζ − λ|.’ • Let f ∈ Hol(Ω), and let λ ∈ ∂Ω be an isolated point of the boundary ∂Ω. Then f is extendable at the point λ if and only if sup0 N and every z ∈ K and such that the  (numerical) product k>N fk (z) converges uniformly with respect to z ∈ K  (hence, the sequence ( NN converges in the space C(K)). If such a convergence takes place, the result f (z) =

∞  k=1

fk (z) =

N  k=1

fk (z) ·



fk (z)

k>N

is a holomorphic function on Ω. The set of the zeros of f is the union of the zeros of the fk , k = 1, 2, . . . .

https://doi.org/10.1017/9781316882108.010 Published online by Cambridge University Press

Appendix D Key Notions of Banach spaces

In this chapter, every vector space is over the field C of complex numbers.

D.1 Normed Spaces and Banach Spaces Let X be a vector space. A function x −→ x on X is called a norm if it satisfies: (i) x + y ≤ x + y for every x, y ∈ X ( ·  is subadditive), (ii) λx = |λ| · x for every x ∈ X and any λ ∈ C, (iii) x = 0 ⇔ x = 0. If  ·  is a norm, ρ(x, y) = x − y is a distance on X (associated with  · ). X equipped with a norm (and with the associated distance) is said to be a normed space. If X is complete as a metric space, X is called a Banach space. In what follows, X denotes a normed space, equipped with a norm  ·  =  · X . space is complete if and only if every absolutely convergent • A normed   series k≥0 xk (xk ∈ X) (i.e. k≥0 xk  < ∞) converges in X (i.e. there exists n x ∈ X such that limn x − k=0 xk  = 0).

D.2 The Baire Category Theorem Every Banach space X (and moreover every complete metric space) is of Baire second category (i.e. for every sequence (Xn )n≥1 , Xn ⊂ X, of closed subsets  with empty interior, we have X  n≥1 Xn ; the subsets that are unions of this last type are said to be of Baire first category). 251 https://doi.org/10.1017/9781316882108.012 Published online by Cambridge University Press

252

Key Notions of Banach spaces

D.3 Duality For a normed space X, we denote X ∗ its dual space, i.e. the space of bounded linear functionals ϕ equipped with the norm ϕ = sup{|ϕ(x)| : x ∈ X, x ≤ 1}. X ∗ is always a Banach space. For reasons of symmetry, we also use the notation x, ϕ = ϕ(x)

(x ∈ X, ϕ ∈ X ∗ ).

• The Hahn–Banach Theorem (1932) (i) Let E ⊂ X be a vector subspace and ϕ0 ∈ E ∗ . Then there exists ϕ ∈ X ∗ such that ϕ|E = ϕ0 and ϕ = ϕ0 . (ii) Let E ⊂ X be a vector subspace and x ∈ X. Then, for every functional ϕ ∈ X ∗ such that ϕ|E = 0 we have |x, ϕ| ≤ distX (x, E) · ϕ. If x  E then there exists a functional ϕ ∈ X ∗ such that ϕ|E = 0 and 1 = |x, ϕ| = distX (x, E) · ϕ. ∗ • Corollary. For every x ∈ X, x = sup{|x, ϕ| : ϕ ∈ X , ϕ ≤ 1}.

• Corollary. Let A ⊂ X, x ∈ X. The following assertions are equivalent: (i) x ∈ spanX (A). (ii) ∀ϕ ∈ X ∗ , ϕ|A = 0 ⇒ ϕ(x) = 0. ∗ • Corollary. Let ϕ ∈ X , E ⊂ X be a vector subspace and ⊥ ∗ E = {ψ ∈ X : ψ|E = 0} (the polar subspace of E). Then

ϕ|E = distX ∗ (ϕ, E ⊥ ). ∗ • Weak topologies. A base of the weak topology σ(X, X ) is defined by

{x ∈ X : |x − x0 , ϕ j | < , j = 1, . . . , n} where n ∈ N, > 0, ϕ j ∈ X ∗ , x0 ∈ X. A base of the weak-star topology σ(X ∗ , X) is defined by {ϕ ∈ X ∗ : |x j , ϕ − ϕ0 | < , j = 1, . . . , n} where n ∈ N, > 0, x j ∈ X, ϕ0 ∈ X ∗ . • Weak-star convergence. Let X be a Banach space and A ⊂ X such that spanX (A) = X. Then a countable sequence (ϕk ) converges σ(X ∗ , X) to 0 if and only if supk ϕk  < ∞ and limk ϕk (x) = 0 ∀x ∈ A. If X is separable, then the unit ball {ϕ ∈ X ∗ : ϕ ≤ 1} is σ(X ∗ , X)-compact. ∗ • Reflexivity. For every x ∈ X, the formula j(x)ϕ = x, ϕ, ϕ ∈ X , defines a functional j(x) ∈ (X ∗ )∗ such that  j(x) = x. X is said to be reflexive if j(X) = (X ∗ )∗ .

https://doi.org/10.1017/9781316882108.012 Published online by Cambridge University Press

D.5 Schauder Bases (1927)

253

• A Banach space X is reflexive if and only if the ball {x ∈ X : x ≤ 1} is σ(X, X ∗ )-compact.

D.4 Examples of Duality 

For 1 ≤ p < ∞, (L p (Ω, μ))∗ = L p (Ω, μ), 1/p + 1/p = 1, with respect to the (bilinear) form realizing the duality   f g dμ, f ∈ L p (Ω, μ), g ∈ L p (Ω, μ).  f, g = Ω

Hence L , with 1 < p < ∞, is reflexive. If supp(μ) is infinite, neither L1 nor L∞ are reflexive. p

∗ • If K is compact, then (C(K)) = M(K), while if Ω is locally compact then ∗ (C0 (Ω)) = M(Ω), with respect to the dualities   f, μ = f dμ, f ∈ C(K), μ ∈ M(K), K

and the analog for C0 (Ω) (see § A.4).

D.5 Schauder Bases (1927) A sequence (ek )k≥1 is called a Schauder basis of a space X if ∀x ∈ X, ∃ a unique sequence (ak ), ak ∈ C such that n

lim x − ak ek = 0. n

The sums Pn x =

n k=1

k=1

ak ek are called the partial sums.

• In a Banach space X, a sequence (ek )k≥1 is a Schauder basis if and only if (i) spanX (ek : k ≥ 1) = X,   (ii) the projections Pn ( k ak ek ) := nk=1 ak ek are well-defined and continuous on Lin(ek : k ≥ 1), and (iii) supn Pn  < ∞. • Remark. There exist separable Banach spaces (and even subspaces of l p = l p (N), p  2) without a Schauder basis (Enflo, 1972).

https://doi.org/10.1017/9781316882108.012 Published online by Cambridge University Press

Appendix E Key Notions of Linear Operators

E.1 Bounded Operators Let X, Y be normed spaces, and let T : X → Y be a linear mapping. The following assertions are equivalent. (i) T is continuous. (ii) T is continuous at the point 0. (iii) T is bounded: ∃ a constant C > 0 such that ∀x ∈ X, T x = T xY ≤ Cx = CxX . (iv) T  = T Op := sup{T x : x ∈ X, x ≤ 1} < ∞ (the best constant C of (iii)). The set of bounded linear operators X → Y, denoted L(X, Y) (and with L(X) = L(X, X)), is a normed space (with the norm  · Op ), complete if Y is complete.  ∗ ∗  • Adjoint operator. If T ∈ L(X, Y) and y ∈ Y , the functional T y is defined ∗   by the requirement x, T y  = T x, y , ∀x ∈ X. The mapping y −→ T ∗ y is linear and bounded, hence T ∗ ∈ L(Y ∗ , X ∗ ); T ∗ is said to be the adjoint operator of T . Clearly T ∗  = T .

E.2 Three Fundamental Principles (1) Closed graph theorem. Let X, Y be Banach spaces, and let T : X → Y be a linear mapping. The following assertions are equivalent. (i) T is continuous. (ii) The graph G(T ) = {(x, y) ∈ X × Y : y = T x} is closed. (iii) limn xn X = 0 and limn T xn − yY = 0 implies y = 0. 254 https://doi.org/10.1017/9781316882108.013 Published online by Cambridge University Press

E.3 The Spectrum

255

(2) The Banach–Steinhaus theorem (1927, principle of equicontinuity, or of uniform boundedness). Let X, Y be Banach spaces, A ⊂ X such that spanX (A) = X and let T n ∈ L(X, Y) be a sequence of continuous linear mappings. The following assertions are equivalent. (i) For every x ∈ X, there exists limn T n x := T x. (ii) supn T n  < ∞, and for every x ∈ A the limit limn T n x exists. The limit T in (i) is always bounded. (3) Open Mapping Theorem (Banach and Schauder, 1932). Let X, Y be Banach spaces and T ∈ L(X, Y). The following assertions are equivalent. (i) The image T (G) of any open set G ⊂ X is open (we say “T is open”). (ii) T X = Y. (iii) There exists a constant c > 0 such that for every y ∈ Y ∗ we have T ∗ y X ∗ ≥ cy Y ∗ . If (i)–(iii) hold, then ∀y ∈ Y, ∃x ∈ X such that T x = y and x ≤ 1c y. • Remark. For a bijective operator T the equivalence of properties (i) and (iii) is obvious (with c = 1/T −1 ); the proof of (3) uses the quotient operator T : X/ Ker(T ) → Y, already bijective. • Corollary. Let X, Y be Banach spaces. (i) If T ∈ L(X, Y) is bijective, it is a homeomorphism.  (ii) If E, F ⊂ X are closed subspaces such that E F = {0}, E + F = X, then the projection PEF (x + y) = x (x ∈ E, y ∈ F) is bounded. (iii) If T : X → Y is linear and continuous for a separable topology τ on Y, then T ∈ L(X, Y) (for example, it could be that τ = σ(Y, Y ∗ )). (iv) The Riemann–Lebesgue lemma. If T n f = fˆ(n), f ∈ L1 (T), then T n  = 1 and limn T n ( f ) = 0 for every trigonometric polynomial f . By (2), limn T n ( f ) = 0 for every f ∈ L1 (T). Remark Other corollaries similar to (iv) can easily be produced.

E.3 The Spectrum Let A be a Banach space equipped with a multiplication operation (x, y) −→ x · y = xy which transforms A into an algebra with unit e ∈ A satisfying xy ≤ x · y (for every x, y ∈ A) and e = 1. Such an algebra A is called a Banach algebra.

https://doi.org/10.1017/9781316882108.013 Published online by Cambridge University Press

256

Key Notions of Linear Operators

∞ • Examples. A = L(X) where X is a Banach space; A = C(K) or A = L (Ω, μ) (equipped with the norm  · ∞ ).

Let A−1 denote the set of invertible elements of A. The spectrum σ(a) = σA (a) of an element a ∈ A is defined by  σ(a) = λ ∈ C : λe − a  A−1 . • Immediate properties of the spectrum. Let A be a Banach algebra with unit e. (i) For every a ∈ A, σ(a) is a non-empty compact set. (ii) The spectral radius r(a) := max{|λ| : λ ∈ σ(a)} coincides with limn an 1/n (Gelfand’s formula). (iii) For A = L(X), the point spectrum of an operator T (the eigenvalues of T ) σ p (T ) = {λ ∈ C : Ker(λI − T )  {0}} is contained in σ(T ). (iv) For A = L(X) and for a bilinear duality between X and X ∗ , we have σ(T ) = σ(T ∗ ). (v) Spectral mapping theorem. For every polynomial f , σ( f (T )) = f (σ(T )).

E.4 Invariant Subspaces Let X be a Banach space, E ⊂ X a closed subspace, and T ∈ L(X). E is said to be an invariant subspace for T if x ∈ E ⇒ T x ∈ E (in brief, T E ⊂ E). The set of all invariant subspaces is denoted Lat(T ). ⊥ ∗ • E ∈ Lat(T ) ⇔ E ∈ Lat(T ). • Lat(T ) is a lattice with respect to the set operations ∪, ∩.

Remark There exists a T ∈ L(l1 ) with the trivial lattice Lat(T ) = {{0}, l1 } (Read, 1984, inspired by Enflo, 1976) and there exist Banach spaces X where Lat(T )  {{0}, X}, ∀T ∈ L(X) (Argyros–Haydon, 2009). For a Hilbert space, the question of the existence of a T with trivial Lat(T ) remains open.

E.5 In a Hilbert Space: Self-adjoint, Unitary, Normal Operators Let H, K be Hilbert spaces and T ∈ L(H, K). We define T ∗ : K → H by (T x, y)K = (x, T ∗ y)H (for every x ∈ H, y ∈ K), which gives a complex

https://doi.org/10.1017/9781316882108.013 Published online by Cambridge University Press

E.5 In a Hilbert Space: Self-adjoint, Unitary, Normal Operators

257

conjugate for certain properties of T ∗ . For example, if T ∈ L(H), then σ(T ∗ ) = σ(T )∗ = {λ : λ ∈ σ(T )}. ∗ • An operator T ∈ L(H) is said to be self-adjoint if T = T , unitary if ∗ ∗ ∗ T T = T T = id (a modification for T ∈ L(H, K) : T T = idH , T T ∗ = idK ), and normal if T T ∗ = T ∗ T .

• Operators A ∈ L(H) and B ∈ L(K) are said to be unitarily equivalent if there exists a unitary operator U : H → K such that UA = BU. • Spectral theorem for a normal operator with a simple spectrum (von Neumann, 1929). Let T ∈ L(H) be a normal cyclic operator (said to be “with simple spectrum”). Then there exists a Borel measure μ on C, with compact support, such that T is unitarily equivalent to the multiplication operator Mz : L2 (μ) → L2 (μ),

Mz f = z f.

We have σ(T ) = supp(μ), and the equivalence class of μ (i.e. {ν ≥ 0 : ν  μ, μ  ν}) is uniquely defined by T (μ is called the scalar spectral measure of T ). • Outline of the proof. (1) We first show that for any normal operator N, we have N = r(N) (spectral radius). (2) By using the spectral mapping theorem (§ E.3), we deduce that for every polynomial in z and z,  f (T ) = r( f (T )) = max{| f (λ)| : λ ∈ σ(T )}, and hence f −→ ( f (T )x, y) is a continuous linear functional on C(σ(T )) (for self-adjoint and/or unitary operators, this step is much simpler than in the general case). (3) We select a cyclic vector x ∈ H, H = spanH (T n x : n = 0, 1, . . . ), and observe (by the Riesz representation theorem, § A.4) that there exists a measure μ ≥ 0 such that  f g dμ ( f (T )x, g(T )x) = σ(T )

for any polynomials f = f (z, z) and g = g(z, z). (4) Setting U( f (T )x) = f , U : H → L2 (μ), we obtain the result.



• Polar decomposition. For every T ∈ L(H, K) (where H and K are Hilbert spaces) such that dim Ker(T ) = dim Ker(T ∗ ), there exists a unitary operator U : H → K such that T = U|T |, where |T | := (T ∗ T )1/2 ≥ 0 is the modulus of T , |T | ∈ L(H).

https://doi.org/10.1017/9781316882108.013 Published online by Cambridge University Press

258

Key Notions of Linear Operators

∗ ∗ ∗ • Corollary. If Ker(T ) = {0}, Ker T = {0}, then the operators T T and T T ∗ ∗ ∗ are unitarily equivalent. (Indeed, T T = U(T T )U .)

• Reducing subspaces of an operator T ∈ L(H): these are the elements of Lat(T ) ∩ Lat(T ∗ ). We have E ∈ Lat(T ) ∩ Lat(T ∗ ) ⇔ E, E ⊥ ∈ Lat(T ) (hence H = E ⊕ E ⊥ where the two subspaces E, E ⊥ are T -invariants). As the closed linear span of a family of reducing subspaces is again in Lat(T ) ∩ Lat(T ∗ ), we deduce that ∀E ∈ Lat(T ) we have E = E  ⊕ E  , where E  ∈ Lat(T ) ∩ Lat(T ∗ ) and E  ∈ Lat(T ) but does not contain any T -reducing subspace (E  is said to be completely non-reducing).

https://doi.org/10.1017/9781316882108.013 Published online by Cambridge University Press

References

A N. I. Akhiezer (1956), On the weighted approximation of continuous functions by polynomials on the real axis. Uspekhi Mat. Nauk 11:4 (70), 3–43. English translation: Amer. Math. Soc. Transl. (2) 22 (1962), 95–137. N. I. Akhiezer (1965), Lectures on Approximation Theory (in Russian), second edition. Nauka, Moscow. English translation: Approximation Theory, Dover, New York (1992).

B L. B´aez-Duarte (2003), A strengthening of the Nyman–Beurling criterion for the Riemann hypothesis. Rend. Lincei (9) Mat. Appl. 14, 5–11. B. Bagchi (2006), On Nyman, Beurling, and Baez-Duarte’s Hilbert space reformulation of the Riemann hypothesis. Proc. Indian Acad. Sci. (Math. Sci.) 116:2, 137–146. M. Balazard (2010), Un si`ecle et demi de recherches sur l’hypoth`ese de Riemann. Gazette des math´ematiciens (Soc. Math. France) 126, 7–24. S. Banach (1932), Th´eorie des op´erations lin´eaires. Monografie Matematyczne, Warsaw. A. Baranov, Yu. Belov and A. Borichev (2013), Hereditary completeness for systems of exponentials and reproducing kernels, Adv. in Math. 235, 525–554. A. Baranov and D. Yakubovich (2016), Completeness and spectral synthesis of nonselfadjoint one-dimensional perturbations of selfadjoint operators, Adv. in Math. 302, 740–798. K. Barbey and H. K¨onig (1977), Abstract Analytic Function Theory and Hardy Algebras. Vol. 593 of Lecture Notes in Mathematics, Springer, Berlin. S. N. Bernstein (1924), Le probl`eme de l’approximation des fonctions continues sur tout l’axe r´eel et l’une de ses applications. Bull. Math. Soc. France 52, 399–410. A. Beurling (1945), On the completeness of {ψ(nt)} on L2 (0, 1). In The Collected Works of Arne Beurling, vol. 2: Harmonic Analysis. Contemporary Mathematicians, Birkh¨auser, Boston (1989), pp. 378–380.

259 Published online by Cambridge University Press

260

References

A. Beurling (1949), On two problems concerning linear transformations in Hilbert space. Acta Math. 81, 79–93. Ia. Blagouchine (2018), The history of the ζ functional equation, and the role of different mathematicians in its proof, A seminar talk at POMI seminar on the history of mathematics, March 1, 2018, www.mathnet.ru/php/conference.phtml?option lang=rus&eventID=10&confid=504. W. Blaschke (1915), Eine Erweiterung des Satzes von Vitali u¨ ber Folgen analytischer Funktionen. S.-B. S¨achs Akad. Wiss. Leipzig Math-Natur. Kl. 67, 194–200. R. P. Boas (1954), Entire Functions. Academic Press, New York. ¨ H. Bohr (1913), Uber die Bedeutung der Potenzreihen unendlich vieler Variablen in der  Theorie der Dirichletschen Reihen anns . Nachr. Ges. Wiss. G¨ottingen. Math.-Phys. Kl. A9, 441–488. A. Borichev (2001), On the closure of polynomials in weighted spaces of functions on the real line. Indiana Univ. Math. J. 50, 829–846. A. Borichev and M. Sodin (2001), Krein’s entire functions and Bernstein approximation problem. Illinois J. Math. 45:1, 167–185. D. G. Bourgin (1946), A class of sequences of functions. Trans. Amer. Math. Soc. 60, 478–518. P. L. Butzer (1983), A survey of the Whittaker–Shannon sampling theorem and some of its extensions. J. Math. Res. Exposition 3, 185–212. P. L. Butzer, P. J. S. G. Ferreira, J. R. Higgins, S. Saitoh, G. Schmeisser, R. L. Stens (2011), Interpolation and Sampling: E. T. Whittaker, K. Ogura and Their Followers. J. Fourier Analysis Appl. 17:2, 320–354. P. L. Butzer, J. R. Higgins, and R. L. Stens (2000), Sampling theory of signal analysis 1950–1995. In Development of Mathematics 1950–2000 (ed. J.-P. Pier), Birkh¨auser, Basel, pp. 193–234.

C A. P. Calder´on (1950), On theorems of M. Riesz and A. Zygmund. Proc. Amer. Math. Soc. 1, 533–535. L. Carleson (1956), Representations of continuous functions. Math. Zeit. 66, 447–451. J. B. Conrey (2003), The Riemann hypothesis. Notices Amer. Math. Soc. March 2003, 341–353. M. Cotlar and C. Sadosky (1979), On the Helson–Szeg˝o theorem and a related class of modified Toeplitz kernels. In Harmonic Analysis in Euclidean Spaces, part 1 (ed. G. Weiss and S. Wainger), vol. 35 of Proceedings of Symposia in Pure Mathematics, American Mathematical Society, Providence, RI, pp. 387–407.

D L. de Branges (1959), The Bernstein problem. Proc. Amer. Math. Soc. 10, 825–832. A. Devinatz and M. Shinbrot (1969), General Wiener–Hopf operators. Trans. Amer. Math. Soc. 145, 467–494. R. A. DeVore and G. G. Lorentz (1993), Constructive Approximation. Springer.

Published online by Cambridge University Press

References

261

J. Duoandikoetxea (2001), Fourier Analysis. American Mathematical Society, Providence, RI. P. L. Duren (1970), Theory of H p Spaces. Academic Press, New York.

F P. Fatou (1906), S´erie trigonom´etriques et s´eries de Taylor. Acta Math. 30, 335–400. ¨ L. Fej´er and F. Riesz (1921), Uber einige funktionentheoretische Ungleichungen. Math. Zeit. 11, 305–314. K. Ford (2002), Vinogradov’s integral and bounds for the Riemann zeta function. Proc. London Math. Soc. (3) 85, 565–633. K. O. Friedrichs (1937), On certain inequalities and characteristic value problems for analytic functions and for functions of two variables. Trans. Amer. Math. Soc. 41, 321–364.

G T. W. Gamelin (1969), Uniform Algebras. Prentice Hall, Englewood Cliffs, New Jersey. F. R. Gantmacher (1966), The Theory of Matrices (in Russian), second edition. Nauka, Moscow. English translation: Chelsea, New York (1960). J. B. Garnett (1981), Bounded Analytic Functions. Academic Press, New York. A. O. Gelfond (1958), Die Rolle der Arbeiten L. Eulers f¨ur die Entwicklung der Zahlentheorie (in Russian, with a summary in German). In Leonhard Euler (zu 250. Geburtstages) (ed. M. Lavrentiev, A. Yushkevich, and A. Grigoriyan), Academy of Sciences of the USSR, Moscow, pp. 96–129. I. M. Glazman and Y. I. Lyubich (1969), Finite-dimensional Linear Analysis (in Russian). Nauka, Moscow. English translation: Finite-dimensional Linear Analysis: A Systematic Presentation in Problem Form, MIT Press, Cambridge, MA (1974). G. Golub and C. Van Loan (1996), Matrix Computations, third edition. Johns Hopkins University Press, Baltimore and London. G. M. Goluzin (1966), Geometric Theory of Functions of a Complex Variable (in Russian). Nauka, Moscow. English translation: American Mathematical Society, Providence, RI (1969). B. Green and T. Tao (2008), The primes contain arbitrarily long arithmetic progressions. Ann. of Math. 167:2, 481–547.

H G. H. Hardy (1913), A theorem concerning Taylor’s series. Quart. J. Pure Math. 44, 147–160. G. H. Hardy (1915), On the mean value of the modulus of an analytic function. Proc. London Math. Soc. (2) 14, 269–277. G. H. Hardy (1922), On the integration of Fourier series. Messenger of Math. 51, 186–192.

Published online by Cambridge University Press

262

References

G. H. Hardy (1941), Notes on special system of orthogonal functions (IV): The orthogonal functions of Whittaker’s cardinal series. Proc. Cambridge Phil. Soc. 37, 331–348. G. H. Hardy and J. E. Littlewood (1916), Some problems of Diophantine approximation: a remarkable trigonometrical series. Proc. Nat. Acad. USA 2, 583–586. G. H. Hardy and J. E. Littlewood (1926), Some new properties of Fourier constants. Math. Ann. 97, 159–209. G. H. Hardy and E. M. Wright (1938), An Introduction to the Theory of Numbers. Sixth edition, Oxford University Press (2008). V. Havin and B. J¨oricke (1994), The Uncertainty Principle in Harmonic Analysis. Springer. H. Hedenmalm, P. Lindquist, and K. Seip (1997), A Hilbert space of Dirichlet series and systems of dilated functions in L2 (0, 1). Duke Math. J. 86, 1–37. H. Hedenmalm, P. Lindquist, and K. Seip (1999), Addendum to “A Hilbert space of Dirichlet series and systems of dilated functions in L2 (0, 1)”. Duke Math. J. 99, 175–178. H. Helson (1964), Lectures on Invariant Subspaces. Academic Press, New York. H. Helson and D. Lowdenslager (1961), Invariant subspaces. In Proc. Intern. Symp. Linear Spaces, Jerusalem, Pergamon Press, Oxford, pp. 251–262. H. Helson and D. Sarason (1967), Past and future. Math. Scand. 21, 5–16. H. Helson and G. Szeg˝o (1960), A problem of prediction theory. Ann. Mat. Pura Appl. 51, 107–138. ¨ G. Herglotz (1911), Uber Potenzreihen mit positiven reellen Teil im Einheitskreise. Berichte Verh. Kgl.-s¨achs. Gesellsch. Wiss. Leipzig, Math.-Phys. Kl. 63, 501–511. J. R. Higgins (1985), Five short stories about the cardinal series. Bull. Amer. Math. Soc. 12:1, 45–89. J. R. Higgins (1996), Sampling Theory in Fourier and Signal Analysis: Foundations. Clarendon Press, Oxford, and Oxford University Press, New York. J. R. Higgins and R. L. Stens, editors (1999), Sampling Theory in Fourier and Signal Analysis: Advanced Topics. Clarendon Press, Oxford. D. Hilbert (1909), Wesen und Ziele einer Analysis der unendlich vielen unabh¨angigen Variablen. Rend. Cir. Mat. Palermo 27, 59–74. D. Hilbert (1912), Gr¨undz¨uge einer allgemeinen Theorie der linearen Integralgleichungen. Teubner, Leipzig. K. Hoffman (1962), Banach Spaces of Analytic Functions. Prentice Hall, Englewood Cliffs, New Jersey. B. Hollenbeck and I. Verbitsky (2000), Best constants for the Riesz projection. J. Funct. Analysis 175, 370–392. R. Hunt, B. Muckenhoupt, and R. L. Wheeden (1973), Weighted norm inequalities for the conjugate function and Hilbert transform. Trans. Amer. Math. Soc. 176, 227– 251.

I I. A. Ibragimov and Y. A. Rozanov (1970), Gaussian Stochastic Processes (in Russian). Nauka, Moscow. English translation: Springer (1978). A. E. Ingham (1936), A note on Hilbert’s inequality. J. London Math. Soc. 11, 237–240.

Published online by Cambridge University Press

References

263

J J. L. Jensen. 1899), Sur un nouvel et important th´eor`eme de la th´eorie des fonctions. Acta Math. 22, 219–251.

K M. Kac (1966), Can one hear the shape of a drum? Amer. Math. Monthly 73:4(2), 1–23. J. P. Kahane and Y. Katznelson (1971), Sur le comportement radial des fonctions analytiques. C. R. Acad. Sci. Paris Ser. A–B 227, A718–A719. J.-P. Kahane and P. G. Lemari´e-Rieusset (1998), S´eries de Fourier et ondelettes. Cassini, Paris. J.-P. Kahane and R. Salem (1963) Ensembles parfaits et s´eries trigonom´etriques. Hermann, Paris. Y. Katznelson (1976), An Introduction to Harmonic Analysis. Dover, New York. C. E. Kenig (1994), Harmonic Analysis Techniques for Second Order Elliptic Boundary Value Problems. CBMS Conference series no. 83, American Mathematical Society, Providence, RI. A. N. Kolmogorov (1925), Sur les fonctions harmoniques conjugu´ees et les s´eries de Fourier. Fund. Math. 7, 24–29. A. N. Kolmogorov (1941), Stationary sequences in Hilbert space (in Russian). Bull. Moscow Univ. Math. 2:6, 1–40. P. Koosis (1966), Weighted polynomial approximation on arithmetic progressions of intervals or points. Acta Math. 116, 223–277. P. Koosis (1980), Introduction to H p Spaces. Cambridge University Press. V. A. Kotelnikov (1933), On the transmission capacity of “aether” and wire in electrocommunications (in Russian). Izdat. Red. Upr. Svyazi RKKA. English translation: http://ict.open.ac.uk/classics/1.pdf. V. A. Kotelnikov (1956), The Theory of Optimum Noise Immunity. McGraw-Hill (1959). Russian original: Izdat. Radio i Svyaz’, Moscow. V. Y. Kozlov (1948), On the completeness of systems of functions {ϕ(nx)} in the space L2 (0, 2π) (in Russian). Doklady Akad. Nauk SSSR 61, 977–980. V. Y. Kozlov (1950), On the completeness of a system of functions of type {ϕ(nx)} in the space L2 (in Russian). Doklady Akad. Nauk SSSR 73, 441–444.

L E. Landau (1927), Vorlesungen u¨ ber Zahlentheorie, vols 1–3. Hirzel, Leipzig. P. D. Lax and R. S. Phillips (1967), Scattering Theory. Academic Press, New York and London. B. Y. Levin (1956), Distribution of Zeros of Entire Functions (in Russian). GITTL, Moscow. English translation: American Mathematical Society, Providence, RI (1980). N. Levinson (1956), On the closure problems and the zeros of the Riemann zetafunction. Proc. Amer. Math. Soc. 7, 838–845. J. Lindenstrauss and L. Tzafriri (1977), Classical Banach Spaces, vols I (1977) and II (1979). Springer.

Published online by Cambridge University Press

264

References

J. E. Littlewood (1925), On inequalities in the theory of functions. Proc. London Math. Soc. 23, 481–519. J. E. Littlewood (1953), A Mathematician’s Miscellany. Methuen, London. Revised edition, Littlewood’s Miscellany (ed. B. Bollob´as), Cambridge University Press (1986). J. E. Littlewood (1970), The “pits effect” for functions in the unit circle. J. Analyse Math. 23, 237–268.

M T. Makino, (2003), The Mathematician K. Ogura and the “Greater East Asia War”. In Mathematics and War (ed. B. Booß-Bavnbek and J. Høyrup), Springer, pp. 326–335. P. Masani (1966), Wiener’s contribution to generalized harmonic analysis, prediction theory and filter theory. Bull. Amer. Math. Soc. 72:1(2), 73–125. C. A. McCarthy and J. Schwartz (1965), On the norm of a finite Boolean algebra of projections and applications to theorems of Kreiss and Morton. Comm. Pure Appl. Math. 18, 191–201. O. C. McGehee, L. Pigno, and B. Smith (1981), Hardy’s inequality and the L1 -norm of exponential sums. Ann. of Math. 113, 613–618. S. N. Mergelyan (1956), Weighted approximation by polynomials (in Russian). Uspekhi Mtem. Nauk 11:5, 107–152. English translation: AMS Transl. Ser. 2 10 (1958), 59–106. Y. Meyer (1992), Wavelets and Operators. Cambridge University Press.

N Z. Nehari (1957), On bounded bilinear forms. Ann. of Math. 65, 153–162. J. H. Neuwirth, J. Ginsberg, and D. J. Newman (1970), Approximation by f (kx). J. Funct. Anal. 5, 194–203. J. H. Neuwirth and D. J. Newman (1967), Positive H 1/2 functions are constant. Proc. Amer. Math. Soc. 18, 958. ¨ F. Nevanlinna and R. Nevanlinna (1922), Uber die Eigenschaften analytischer Functionen in der Umgebung einer singul¨aren Stelle oder Linie. Acta Soc. Sci. Fenn. 50:5, 1–46. N. Nikolski (1980), Lekzii ob Operatore Sdviga (in Russian). Nauka, Moscow. N. Nikolski (1986), Treatise on the Shift Operator. Springer. N. Nikolski (1995), Distance formulae and invariant subspaces, with an application to localization of zeros of the Riemann ζ-function. Ann. Inst. Fourier 45:1, 143–159. N. Nikolski (2002), Operators, Functions, and Systems, vols 1 and 2. American Mathematical Society, Providence, RI. N. Nikolski (2012), In a shadow of the RH: cyclic vectors of Hardy spaces on the Hilbert multidisc. Ann. Inst. Fourier 62:5, 1601–1626. N. Nikolski and A. Volberg (1990), Tangential and approximate free interpolation. In Analysis and Partial Differential Equations (ed. C. Sadosky), Marcel Dekker, New York, pp. 277–299.

Published online by Cambridge University Press

References

265

B. Nyman (1950), On the one-dimensional translation group and semi-group in certain function spaces. Thesis, Uppsala University.

O K. Ogura (1920), On a certain transcendental integral function in the theory of interpolation. Tˆohoku Math. J. 17, 64–72. B. K. Øksendal (1971), A short proof of the F. and M. Riesz theorem. Proc. Amer. Math. Soc. 30, 204.

P R. E. A. C. Paley and N. Wiener (1934), Fourier Transforms in the Complex Domain. Vol. 19 of American Mathematical Society Colloquium Publications, Providence, RI. A. Papoulis (1984), Signal Analysis. McGraw-Hill. M. Pavlovi´c (2004), Introduction to Function Spaces on the Disk. Matematiˇcki Institut SANU, Belgrade. V. V. Peller (2003), Hankel Operators and their Applications. Springer. V. V. Peller and S. V. Khruschev (S. V. Hruschev) (1982), Hankel operators, best approximations and stationary Gaussian processes (in Russian). Uspekhi Mat. Nauk 37:1, 53–124. English translation: Russian Math. Surveys 37:1 (1982), 61–144. R. P´erez-Marco (2011), Notes on the Riemann hypothesis. In Jornadas sobre los problemas del milenio, Barcelona 1–3 junio, 2011. E. Phragm´en and E. Lindel¨of (1908), Sur une extension d’un principe classique de l’analyse. Acta Math. 31, 381–406. ¨ A. I. Plessner (1927), Uber das Verhalten analytischer Funktionen am Rande ihres Definitionsbereichs. J. Reine Angew. Math. 158, 219–227. G. P´olya and G. Szeg˝o (1925), Aufgaben und Lehrs¨atze aus der Analysis, vols 1, 2. Springer, Berlin. English translation: Springer (1972). S. C. Power (1982), Hankel Operators on Hilbert Space. Vol. 64 of Pitman Research Notes in Mathematics, Pitman. I. I. Privalov (1941), Boundary Properties of Analytic Functions (in Russian). Moscow (second edition 1950). German translation: Deutscher Verlag, Berlin (1956).

R C. Reid (1970), Hilbert. Springer, New York. ¨ F. Riesz (1923), Uber die Randwerte einer analytische Funktion. Math. Z. 18, 87–95. ¨ F. Riesz and M. Riesz (1916), Uber die Randwerte einer analytische Funktion. In Quatri`eme Congr`es des Math. Scand., Stockholm, pp. 27–44. F. Riesz and B. Sz˝okefalvi-Nagy (1955), Lec¸ons d’analyse fonctionnelle. Akad´emiai Kiado, Szeged.

Published online by Cambridge University Press

266

References

M. Riesz (1927), Sur les fonctions conjugu´ees. Math. Zeit. 27, 218–244. M. Rosenblum (1962), Summability of Fourier series in L p (μ). Trans. Amer. Math. Soc. 105:1, 32–42. Y. A. Rozanov (1963), Stationary Stochastic Processes (in Russian). Fizmatgiz, Moscow. English translation: Holden-Day, San Francisco (1967). W. Rudin (1956), Boundary values of continuous analytic functions. Proc. Amer. Math. Soc. 7, 808–811. W. Rudin (1962), Fourier Analysis on Groups. Wiley, New York. W. Rudin (1998), Analyse r´eelle et complexe, third edition. Dunod, Paris.

S K. Sabbagh (2002), The Riemann Hypothesis: The Greatest Unsolved Problem in Mathematics. Farrar, Straus and Giroux, New York. R. Salem (1953), Sur une proposition e´ quivalente a` l’hypoth`ese de Riemann. C. R. Acad. Sci. Paris 236, 1127–1128. D. Sarason (1994), Sub-Hardy Hilbert Spaces in the Unit Disk. University of Arkansas Lecture Notes, no. 10, Wiley, New York. C. E. Shannon (1948), A mathematical theory of communication. Bell System Technical Journal 27, July and October, 379–423 and 623–656. C. E. Shannon (1949), Communication theory of secrecy systems. Bell System Technical Journal 28, October, 656–715. C. E. Shannon (1950), Programming a computer for playing chess. Philosophical Magazine (7) 41:314, 256–275. J. H. Shapiro (1993), Composition Operators and Classical Function Theory. Springer, New York. B. Simon (2005), Orthogonal Polynomials on the Unit Circle, Part 1: Classical Theory. American Mathematical Society, Providence, RI. V. I. Smirnov (1928a), Sur la th´eorie des polynˆomes orthogonaux a` une variable complexe. J. Leningrad Fiz.-Mat. Obsch. 2:1, 155–179. V. I. Smirnov (1928b), Sur les valeurs limites des fonctions r´eguli`eres a` l’int´erieur d’un cercle. J. Leningrad Fiz.-Mat. Obsch. 2:2, 22–37. V. I. Smirnov (1932), Sur les formules de Cauchy et Green et quelques probl`emes qui s’y rattachent. Izvestia AN SSSR, ser. fiz.-mat. 3, 338–372. V. I. Smirnov (1988), Œuvres choisies: Analyse complexe et th´eorie de diffusion (in Russian). University of Leningrad. M. N. Spijker, S. Tracogna, and B. Welfert (2003), About the sharpness of the stability estimates in the Kreiss matrix theorem. Math. Comp. 72, 697–713. T. P. Srinivasan (1963), Simply invariant subspaces. Bull. Amer. Math. Soc. 69, 706–709. J. M. Steele (2004), The Cauchy–Schwarz Master Class: An Introduction to the Art of Mathematical Inequalities. Cambridge University Press. E. Stein (1993), Harmonic Analysis: Real-variable Methods, Orthogonality, and Oscillatory Integrals. Princeton University Press, Princeton, New Jersey. G. Szeg˝o (1920), Beitr¨age zur Theorie der Toeplitzsche Formen, I. Math. Zeit. 6:3/4, 167–202.

Published online by Cambridge University Press

References

267

¨ G. Szeg˝o (1921), Uber die Randwerte einer analytischen Funktion. Math. Ann. 84:3/4, 232–244.

T J. E. Thomson (1991), Approximation in the mean by polynomials. Ann. of Math. (2), 133:3, 477–507. E. C. Titchmarsh (1939), The Theory of Functions. Oxford Science Publications. E. C. Titchmarsh (1951), The Theory of the Riemann Zeta-function. Oxford Science Publications. O. D. Tsereteli (1975), Metric properties of conjugate functions (in Russian). Itogi Nauki i Techniki Sovrem. Probl. Mat. 7, 18–57. English translation: J. Soviet Math. 7 (1977), 309–414.

V S. Verblunsky (1936), On positive harmonic functions (second paper). Proc. London Math. Soc. (2) 40, 290–320. H. von Koch (1902), Ueber die Riemann’sche Primzahlfunction. Math. Annalen 55, 441–464.

W H. Weyl (1908), Singul¨are Integralgleichungen. Math. Ann. 66, 273–324. E. T. Whittaker (1915), On the functions which are represented by the expansions of the interpolation theory. Proc. Royal Soc. Edinburgh Ser. A 35, 181–194. E. T. Whittaker (1924), The Calculus of Observations: A Treatise on Numerical Mathematics. Blackie, London. N. Wiener (1930), Generalized harmonic analysis. Acta Math. 55, 117–258. N. Wiener (1932), Tauberian theorems. Ann. of Math. (2) 33, 1–100. N. Wiener (1933), The Fourier Integral and Certain of its Applications. Cambridge University Press, New York. N. Wiener (1949), Extrapolation, Interpolation, and Smoothing of Stationary Time Series: With Engineering Applications. MIT Press, Cambridge, MA, and Wiley, New York. N. Wiener and P. R. Masani (1957), The prediction theory of multivariate stochastic processes, I: The regularity condition. Acta Math. 98, 111–150. N. Wiener and P. R. Masani (1958), The prediction theory of multivariate stochastic processes, II: The linear predictor. Acta Math. 99, 93–137. A. Wintner (1944), Diophantine approximation and Hilbert’s space. Amer. J. Math. 66, 564–578. H. Wold (1938), A Study in the Analysis of Stationary Time Series. Almquist och Wiksell, Uppsala.

Z A. Zygmund (1959), Trigonometric Series, vols I and II. Cambridge University Press.

Published online by Cambridge University Press

Notation

Sets and Measures C - the complex plane T = {z ∈ C : |z| = 1} C+ = {z ∈ C : Im(z) > 0} C+ = {z ∈ C : Re(z) > 0} D(z, r) = {ζ ∈ C : |z − ζ| < r} D = D(0, 1) D∞ 2 - Hilbert multi-disk, § 6.6.5 P - Theorem 1.2.1 Pa - Corollary 1.4.4 Pn - Exercise 5.6.2(f) Lat(T ), Lat(T ) - § 1.1 σ(θ) - spectrum of a function, Definition 3.2.3 m - normalized Lebesgue measure, § 1.3 μa , μ s - § 1.3

Spaces and Operations H 2 = H 2 (T) - Definition 1.3.4 H 2 (T, μ) = H 2 (μ) - Definition 1.5.1 H p = H p (T) - § 1.3.1 H p (T, μ) = H p (μ) - § 2.9 H ∞ = H ∞ (T) - Exercise 1.8.3 W = W(T) - § 5.3.1 H02 - Theorem 1.7.6 H02 (μ) - Lemma 1.6.4 H p (D) - Definition 2.2.1 268 Published online by Cambridge University Press

Notation

269

H ∞ (Ω) - § 3.4.1 Ca (D) - § 5.4.1 Wa (D) - Exercise 5.6.2 Mult(X) - multipliers of X, Exercises 1.8.3(a), 4.9.2, 6.7.1 D - Smirnov class - Definition 3.3.1 D(Ω) - § 3.4.1 N - Nevanlinna class - Definition 3.3.1 N(Ω) - § 3.4.1 E f - invariant subspace generated by f , Corollary 1.4.4 Lin(A) - linear hull of A spanX (A) = span(A) - closed linear hull of A, § 1.1 closX (A) = clos(A) - the closure (the adherence) of A, § 1.3.1, Corollary 1.4.4 Functions, Constants, and Transforms χA - characteristic function of A (if x ∈ A, χA (x) = 1, otherwise χA (x) = 0) fin , fout - Theorem 1.7.2 Vμ - singular function with measure μ, Corollary 2.6.4 Γu - Herglotz transform of u, Exercise 2.8.4(c) Γ - Euler gamma function, Theorem 6.1.5 H(u) - Hilbert transform of u, Exercise 2.8.4(d) P+ - Riesz projection, Exercise 2.8.3(g) PE - orthogonal projection on E, Appendix C.2 PLM - skew projection, Definition 4.2.1 A(L, M) = AH (L, M) - the angle between L and M, Definition 4.3.1 F - Fourier transform, Appendix A.6 F∗ - Mellin transform, § 6.3.3 Hϕ - Hankel operator, § 4.7.2 τ s - translation, Definition 5.1.1, § 6.3.2(b) Dt - dilatation (dilation), Lemma 6.2.3 Mz - shift operator, § 1.8.2 ζ(s) - Euler zeta function, Definition 6.1.1 sinc(t) - sinus cardinalis (= sin(πt) πt ), § 5.7 b(X) - basis constant, § 4.1.1(e) ub(X) - unconditional basis constant, § 4.8 w ∈ (HS ) - Helson–Szeg˝o weight, Definition 4.6.2

Published online by Cambridge University Press

Index

Adamyan, Vadym, 147 Ahlfors, Lars, 94 Akhiezer, Naum I., 79 Algebra disk, 171 Wiener, 157 Wiener Wa = W+ , 158 Wiener (analytic), 171 Amp`ere, Andr´e-Marie, 101 Amplitude distortion, 157 Angle between two subspaces, 115 Approximate identity, 242 Arago, Franc¸ois, 101 Argyros, Spiros, 256 Arnold, Vladimir, 22 Arov, Damir, 147 Artin, Emil, 48 Arveson, William, 1 Asymptotics of reproducing kernels, 224 Babenko, Ivan K., 106, 120, 125 Backward shift, 74 B´aez-Duarte, Luis, 229, 230 Bagchi, Bhaskar, 230 Baire, Ren´e-Louis, 235 Balazard, Michel, 229 Banach algebra, 255 space, 251 Banach, Katarzyna, 111 Banach, Stefan, 111, 113, 235, 238, 255 Baranov, Anton, xv Barbey, Klaus, 35 Basis finite, 130 orthonormal, 249

Schauder, 70, 106, 113, 141, 149, 253 summation, 148 unconditional, 129 Bateman, Harry, 177 Bernoulli, Johann, 188 Bernstein, Sergei N., 68, 79 Beurling, Arne, 3, 11, 14, 35, 78, 79, 211, 217, 231 Bianchi, Luigi, 48 Biorthogonal pair, 109 Biot, Jean-Baptiste, 107 Birkhoff, George David, 7, 213 Blaschke condition, 46 product of f , 49 Blaschke, Wilhelm, 46, 48 Blumenthal, Otto, 68 Bohr, Harald, 68, 215 Bohr, Niels, 212 Bombieri, Enrico, 228 ´ Borel, Emile, 178, 182, 235 Borichev, Alexander, 79 Born, Max, 68 Bourgin, David, 231 Burkholder, Donald, 146 Butzer, Paul Leo, 178 Calder´on, Alberto, 81 Carath´eodory, Constantin, 245 Carleson, Lennart, 1, 3, 14, 73, 162 Carroll, Lewis, 197 Catherine the Great, 188 Cauchy, Augustin-Louis, 1, 178 Champollion, Jean-Franc¸ois, 107 Chapman–Kolmogorov, 22 ´ Charpentier, Eric, xv

270 https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

Index

Chebyshev, Pafnuty, 26 Chowla, Sarvadaman, 77 Class D, 104 Nevanlinna, 93, 104 Smirnov, 93 Collingwood, Edward, 77 Completeness of the characters, 206 of the dilations, 206 of the polynomials, 78 of the translations, 206 Complex derivative, 243 Condition Muckenhoupt (A2 ), 146 Condorcet, Nicolas, 189 Conrey, John Brian, 228 Constant basis, 106, 110 unconditional basis, 106, 130 uniform minimality, 110 Convergence of an orthogonal series, 248 unconditional, 248 Convolution in L p (G), 242 Cotlar, Mischa, 147 Counillon, C., xv Courant, Richard, 68 Cram´er, Harald, 18 Criterion of (Dn )-cyclicity, 232 Crocodile, Littlewood’s, 174 Cyclic functions of H p , 63 Cyclic vector, 58 Cyclicity of polynomials, 220 Davenport, Harold, 77 De Branges, Louis, 1, 3, 79 De la Vall´ee-Poussin, Charles, 178 Decomposition Lebesgue, 238 Wold–Kolmogorov, 29, 36 Dedekind, Richard, 197 Denjoy, Arnaud, 235 Density of the polynomials, 240 Descartes, R´en´e, 188 Description of the spaces H 2 (μ), 60 Diderot, Denis, 189 Dieudonn´e, Jean, 11 Differential form closed, 244 exact, 244

271

Dilatation (dilation) Dt , 199 orthogonal, 223 Dirac delta, 240 Dirichlet, Gustav Lejeune, 197 Distance function, 209 Dodgson, Charles, 197 Domain frequency, 152 Jordan, 77, 96, 104 Smirnov, 104 spectral, 152 time, 152 Domar, Yngve, 14 Douglas, Ronald, 1, 3 Du Bois-Reymond, Paul, 2, 171, 173 Duality, 252 Duren, Peter Larkin, 35 Eddington, Arthur, 177 Ehrenpreis, Leon, 64, 65 Eigenvalue, 256 Einstein, Albert, 3, 4 Energy density at the frequency λ, 157 of a signal, 151 Enflo, Per, 253, 256 Equation Cauchy–Riemann (C-R), 81, 243 Erd˝os, Paul, 213 Esseen, Carl-Gustav, 14 Euclid, 190 Euler infinite product, 191 Euler zeta (or ζ) function, 190 Euler, Leonhard, 145, 187–189, 227, 228 Evaluation functional, 209 Factorization Smirnov canonical, 53 Farey, John, 229 Fatou, Pierre, 1, 49, 235 Fefferman, Charles, 1, 3 Fej´er, Lip´ot, 28, 79 Filter all-pass, 157 band-pass, 157 causal, 157 correctly observable, 151 finite-power, 151 finite-power stationary, 151 ideal band-pass, 157 identification, 167, 169, 170

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

272

Filter (cont.) inverse problem, 167 linear, 151 phase correction, 157 recursive, 166 stable, 157 stable stationary, 151 stationary, 151 Foias, Ciprian, 1, 3 Formula Cauchy, 104, 244 Gelfand, 256 Green, 104 Jensen, 43 Szeg˝o–Verblunsky–Kolmogorov, 29, 57 transfer, 241 Fourier series, 108 Fourier transform, 241 Fourier, Joseph, 1, 107, 145 Fr´echet, Maurice, 18 Fractional part, 190 Franel, J´er¨ome, 229 Fr´echet, Maurice, 111 Frequency characteristic, 156 Frequency response, 156 Freudenthal, Hans, 196 Friedmann, Alexander, 26 Friedrichs, Kurt Otto, 146 Frostman, Otto, 18 Fuhrmann, Paul, 1, 3 Function (Dn )-cyclic, 212 Mζ -cyclic, 217 T -cyclic, 217 cyclic of L2 (μ), 78 cyclic with respect to the semigroup T , 214 Euler zeta (or Riemann zeta), 187, 190, 227 Green’s, 104, 105 harmonic, 243 inner, 32, 102 inner (in the sense of Beurling), 15 inner in C+ , 200 maximal, 35 outer, 25, 94 singular “at infinity”, 202 singular inner, 54 Functional equation Riemann, 187 Future of a process, 20 Gabriel, Robert Mark, 79, 230 Gamelin, Theodore, 1, 3, 35

Index

Gamow, George, 212 Garnett, John, 1, 3, 35 Gauss, Carl, 189, 190, 196 Gelfand, Israel, 14 Gelfond, Alexander Osipovich, 187, 227, 228 Glazman, Israel, xiv G¨odel, Kurt, 68 Godement, Roger, 11 Goldbach, Christian, 188 Goluzin, Gennadiy, 26 Gordan, Paul, 68 Gram, Jørgen, 136 Greatest common divisor of τ, 86 Greczek, Stefan, 111 Green, Ben, 225 Grothendieck, Alexander, 11 Group of dilations, 205 Gsell, Katharina, 189 Haar, Alfred, 68, 235 Hadamard, Jacques, 6, 178, 182 Hall, Tord, 14 Hankel, Hermann, 128 Hardy space, 4 H 2 (T), 10 H p (T), 13 H p , 34 abstract, 11 associated with μ, 16 Hardy, G. H., xiii, 1, 2, 4, 79, 173, 177, 183, 194, 202, 213, 228, 235 Harmonic conjugate, 68, 70 Hasse, Helmut, 48 Hausdorff, Felix, 235 Havin, Victor, 1, 3, 26, 35 Haydon, Richard, 256 Hecke, Erich, 48, 68 Heisenberg, Werner, 212 Hellinger, Ernst, 68 Helson set, 162 Helson, Henry, 1, 3, 11, 35, 60, 105, 164, 165 Helson–Szeg˝o (HS), 146 Hereditary completeness, 149 Herglotz, Gustav, 78 Higgins, J. R., 178 Hilbert multi-disk D∞ 2 , 216 space, 247 Hilbert, David, 1, 2, 18, 48, 67, 102, 228 Hille, Einar, 18 HMW (Hunt–Muckenhoupt–Wheeden), 146 Hodge, William, 177

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

Index

H¨older, Irmgard, 213 H¨older, Otto, 213 Hollenbeck, Brian, 148 Holomorphic extension, 246 Holomorphy, 237 H¨ormander, Lars, 18 Hunt, Richard, 146 Identity Parseval, 248 Impulse response, 155, 156 Inequality Cauchy–Schwarz, 240, 247 Fej´er–Riesz, 79 H¨older, 239 Hilbert, 79 Ingham, 80 Jensen, 43, 77, 105 Jensen’s convexity, 240 Kolmogorov weak type, 73 McCarthy–Schwartz, 106, 141 von Neumann, 81 Infimum Szeg˝o, 27, 57 Ingham, Albert, 2, 77, 80, 213 Inner part of f , 57 Integral Lebesgue, 234 Integral representations of the function ζ, 190, 192 Invariant subspaces of H p , 63 of L2 (T), 14 Jensen’s inequality with the harmonic measure, 44 Jensen, Johan, 1, 42, 45 Jordan domain , 245 J¨oricke, Burglind, 35 Kac, Mark, 111, 167, 213 Kahane, Jean-Pierre, 1, 3, 164 Kakutani, Shizuo, 238 Kantorovich, Leonid, 26 Karhunen, Kari, 94 Katznelson, Yitzhak, 73 Kernel Dirichlet, 37 Fej´er, 38 Poisson, 38 reproducing of H 2 (D∞ 2 ), 218, 222 Khrabrov, Andre¨ı, xv Klein, Felix, 48

273

Klein, Oscar, 212 Koch, Elise, 197 Kolmogorov, Andrey N., xiii, 1, 3, 22, 28, 29, 36, 57, 72, 78, 150, 235 Kolmogorov–Arnold–Moser, 22 K¨onig, Hermann, 35 Koosis, Paul, 35, 79 Korobov, Nikolai, 229 Kotelnikov, Vladimir A., 150, 176, 179, 180, 184, 185 Kozlov, V., 227, 231, 232 Kramers, Henrik, 212 Krein, Mark, 147 Kronecker, Leopold, 128 Lacey, Michael, 1 Lagrange, Joseph-Louis, 107 Landau, Edmund, 2 Landau, Lev, 212 Langevin, Paul, 182 Laplace, Pierre-Simon, 107, 145, 189 Lasker, Emanuel, 68 Lattice, 233 Laugwitz, Detlef, 196 Lax, Peter, 1, 3, 204, 230 Least common multiple of τ, 87 Lebedev, Nikolai A., 26 Lebesgue L p (μ) spaces, 239 point, 50, 237 Lebesgue, Henri, xiii, 1, 6, 18, 235 Lehto, Olli, 94 Leibniz, Gottfried, 6 Lemma Fatou, 237 Kolmogorov, 116 Neuwirth–Ginsberg–Newman, 220 Paley–Wiener, 203, 205 Riemann–Lebesgue, 241, 255 Wiener, 158 Levi-Civita, Tullio, 213 Levinson, Norman, 6, 229 Lichtenstein, Leon, 213 Limit non-tangential, 50 Lindel¨of, Ernst Leonard, 93, 98–100, 102, 205, 231 Liouville, Joseph, 101 Littlewood, J. E., 1, 2, 76, 173, 177, 202, 231, 235 Logarithmic residue, 91 Lowdenslager, David, 11, 35 Lozinsky, Sergey, 26

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

274

Luzin, Nikolai, 1, 3, 4, 22, 77, 235 Lyubich, Yuri, xiv M-basis (strong), 149 Majorant Phragm´en–Lindel¨of, 97, 99 Masani, Pesi, 36, 78, 150, 176 Matrix Gram, 106, 135, 149 Laurent, 155 Toeplitz, 155 Matveev, R. F., 36 Maximal function Szeg˝o, 78, 99 Mazur, Stanisław, 111 McCarthy, C. A., 148 McGehee, Oscar, 80 Mean Fej´er, 37 Poisson, 37 Measure, 233 absolutely continuous, 238 spectral of a process, 21 Mellin, Robert Hjalmar, 205 Mergelyan, Sergey, 79 Mittag-Leffler, G¨osta, 205 M¨obius, August Ferdinand, 128 Monge, Gaspard, 107 Montel, Paul, 236 Moving averages, 169, 170 Muckenhoupt, Benjamin, 146 Multiplier, 32 of a family of vectors, 139 Nasar, Sylvia, 64 Nash, John, 64 Nehari, Zeev, 147 Neovius, Otto, 93 Neuwirth, J. H., 64 Nevanlinna, Arne, 94 Nevanlinna, Rolf, 1, 3, 92, 93 Newman, Donald, 64, 81 Newton, Isaac, 188 Nikodym, Otto, 111, 235 Norm, 251 Nyman, Bertil, 14, 190, 229 Nyquist, Harry, 181 Ogura, Kinnosuke, 180–182 Øksendal, Bernt, 18 Operator adjoint, 254 angular between the future and the past, 127

Index

bounded, 254 Hankel, 127 Hilbert, 106, 119, 146 normal, 257 self-adjoint, 257 shift Mz , 30 unitarily equivalent, 257 unitary, 257 Optimal prediction of a state, 126 Orlicz, Władysław, 111 Orthogonal complement, 248 decomposition, 248 projection, 248 Orthogonalization, Gram–Schmidt, 249 Orthogonalizer, 137 Outer function, 33, 58, 82 Beurling, 56 Outer part of f , 57 Paley, Raymond, 1, 3, 202 Parseval, Marc-Antoine, 1 Part integer, 190 Past of a process, 20 Pauli, Wolfgang, 212 Pavlovi´c, Miroslav, 35 Peller, Vladimir, 147 P´erez-Marco, Ricardo, 228 Phase lag at the frequency λ, 157 of x at the frequency λ, 157 Phillips, Ralph, 1, 3 Phragm´en, Lars Edvard, 98–100, 102 Pigno, Louis, 80 Plancherel, Michel, 235 Plessner, Abraham, 78 Poincar´e, Henri, 98, 177 Poisson, Sim´eon Denis, 1, 101, 107, 178 Polar decomposition, 257 Pollak, Aron, 162 P´olya, George, 105, 202 Polynomials Fej´er, 219, 242 orthogonal, 249 Principle compactness (Montel), 245 conformal mappings (Riemann), 245 equicontinuity, 255 generalized maximum, 96, 97 isolated zeros, 245 Littlewood subordination, 74

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

Index

maximum, 245 Phragm´en–Lindel¨of, 96, 99, 105 subordination, 81 uniform boundedness, 255 Privalov, Ivan, 1, 3, 35, 77 Problem Bernstein, 79 direct of filtering, 157 inverse of filtering, 157 of causality, 158 of optimal prediction, 21 Wintner–Beurling, 214 Process regular, 21 singular, 21 strongly regular, 126 Product infinite, 246 Projection Riesz, 67, 68, 70, 174 skew, 106, 114 Radon, Johann, 235 Ramanujan, Srinivasa, 2, 77 Read, Charles, 256 Reid, Constance, 68 Residue logarithmic, 208 Riemann hypothesis (RH), 190, 195, 211 Riemann, Bernhard, 128, 190, 196, 228 Riesz, Frigyes, xiii, 1, 3, 18, 35, 47, 78, 79, 235, 238 Riesz, Marcel, 1–3, 18, 35, 70, 81, 235 Rosenblum, Marvin, 149 Rozanov, Yuri, 36 Rudin, Walter, 81, 162 Runge, Carl, 48 σ-algebra (Borel), 234 Sabbagh, Karl, 228 Sadosky, Cora, 147 Saks, Stanisław, 111, 235 Salem, Rapha¨el, 164, 229 Sampling, 176 Sarason, Donald, 1, 3 Sarnak, Peter, 228 Scalar product, 247 Schauder, Juliusz, 111, 255 Schmidt, Erhard, 68 Schoenfeld, Lowell, 229 Schur, Issai, 1, 3 Schwartz, Laurent, 11

275

Schwartz, Jacob, 131, 148 Semigroup of characters, 201 Sequence dual, 109 minimal, 109 uniformly minimal, 109 Series Fourier, 106, 108 generalized Fourier, 110 Set Helson, 164, 175, 186 independent, 164 Sexton, M., 189 Shannon, Claude, 150, 176, 180, 184 Shapiro, Harold, 64 Sieve of Eratosthenes, 231 Signal, 150 bounded, 151 causal, 157 input harmonic of frequency λ, 152 Signal processing, 147 Sinus cardinal, 179 Smirnov canonical factorization , 55 Smirnov, Vladimir I., 1, 3, 26, 35, 36, 78, 92, 94, 99, 104, 105 Smith, Brent, 80 Snow, C. P., 2 Sobolev, Sergei, 26 Space H ∞ (D∞ 2 ), 219 Bergman, 81 Hardy H 2 (C+ ), 200 Hardy in D∞ 2 , 216 Paley–Wiener, 183 reflexive, 252 Spectral radius, 256 Spectrum, 256 energy, 157 of an inner function, 88, 89, 104 point, 256 Spencer, Donald, 77 Spijker, Marc, 143, 148 Srinivasan, T. P., 13, 60, 61, 79 Stationary process, 20, 147 Stein, Elias, 1, 3, 35 Steinhaus, Hugo, 68, 111, 235 Stens, R. L., 178 Stieltjes, Thomas Joannes, 1 Stolz angle, 50 Str¨omberg, T., 213 Struik, Dirk, 189

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

276

Subspace completely non-reducing, 9 generated by an element, 58 invariant, 5, 256 invariant under dilatation, 206 reducing, 5, 258 Sum partial, 110, 253 Summation method, 148 Sz.-Nagy, B´ela, 1, 3, 18 Szeg˝o maximal functions , 53 Szeg˝o, G´abor, 1, 3, 27–29, 35, 57, 78, 105 Szpilrajn, Edward, 11 Tamarkin, Yakov D., 26 Tao, Terence, 225 Teller, Edward, 212 Theorem Riesz, 13 Whittaker–Ogura–Kotelnikov, 183 Wiener 1/ f , 64 Bagchi, 230 Baire category, 251 Banach, 163–165 Banach–Steinhaus, 255 Beppo Levi, 237 Beurling, 14, 15 Blaschke, 46 boundary uniqueness, 15 boundary uniqueness for H 1 , 44 closed graph, 254 Euler–Riemann, 193 Fatou, 50, 78 Green–Tao, 225 Hahn–Banach, 252 Helson, 11, 62 Helson and Szeg˝o, 120, 126 Herglotz, 54 Kolmogorov, 21 Kronecker Solenoid, 164 Lax, 204 Lebesgue (dominated convergence), 237 Liouville, 98, 100 McCarthy and Schwartz, 131, 148 Nehari, 128 Neuwirth–Ginsberg–Newman, 220 Nyman, 207, 231 open mapping, 255 Paley–Wiener, 183 Plancherel, 202 Pythagoras, 248

Index

Radon–Nikodym, 238 Riesz brothers, 17, 41, 163 Riesz representation, 41, 238, 250 Rouch´e, 245 Rudin and Carleson, 163 sampling, 178 Smirnov, 25, 27, 83 Smirnov–Beurling, 59 spectral, 257 Szeg˝o–Verblunsky–Kolmogorov, 57 Wiener, 61, 153, 154 Thomson, James, 79 Thorin, Olof, 18 Tibbets, Paul, 4 Titchmarsh, Edward, 2, 194, 228 Toeplitz, Otto, 68 Topology weak σ(X, X ∗ ), 252 weak-star σ(X ∗ , X), 252 Tracogna, Stefania, 143, 148 Transfer function, 155, 156 Transform Bohr, 215, 231 Fourier–Plancherel, 241 Herglotz, 69 Mellin, 204 Tsereteli, Otar, 73 Tumarkin, Genrich, 104 Ulam, Stanisław, 111 Uncertainty principle, 35 Variation of a measure, 233 Verbitsky, Igor, 148 Verblunsky, Samuel, 28, 29, 36, 57, 78 Vinogradov, Ivan M., 229 Vinogradov, Stanislav A., 175 Von Koch, Helge, 228 Von Neumann, John, 28, 68, 257 Watson, G. N., 177 Wegener, Alfred, 3, 4 Weierstrass, Karl, 1, 128, 205 Weight Helson–Szeg˝o, 124 Weighted density of polynomials, 59 Weighted variation of the phase, 159 Weil, Andr´e, 213 Welfert, Bruno, 143, 148 Weyl, Hermann, 68, 79 Wheeden, Richard, 146

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

Index

Whittaker, E. T., 176, 177, 180, 181 Wiener’s theorem, 7 Wiener, Norbert, xiii, 1, 3, 6, 7, 36, 78, 150, 176, 186, 202, 213, 229 Wintner, Aurel, 211, 212, 221, 222, 231 Wintner–Beurling problem, 229 Wirtinger, Wilhelm, 48 Wright, Frank Lloyd, 163

277

Yakubovich, Vladimir, 26 Yoccoz, Jean-Christophe, 98 Young, W. H., 67 Zagier, Don, 64 Zermelo, Ernst, 68 Zermelo–Frenkel, 68 Zero divisor, 48 Zero multiplicity function, 48 Zygmund, Antoni, 1, 3, 79, 81, 202, 230

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press

https://doi.org/10.1017/9781316882108.014 Published online by Cambridge University Press